European Robotics Symposium 2006

Springer Tracts in Advanced Robotics Volume 22 Editors: Bruno Siciliano · Oussama Khatib · Frans Groen Springer Tracts...

Author: Henrik Iskov Christensen

21 downloads 1929 Views 18MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Springer Tracts in Advanced Robotics Volume 22 Editors: Bruno Siciliano · Oussama Khatib · Frans Groen

Springer Tracts in Advanced Robotics Edited by B. Siciliano, O. Khatib, and F. Groen

Vol. 21: Ang Jr., H.; Khatib, O. (Eds.) Experimental Robotics IX 618 p. 2006 [3-540-28816-3] Vol. 20: Xu, Y.; Ou, Y. Control of Single Wheel Robots 188 p. 2005 [3-540-28184-3] Vol. 19: Lefebvre, T.; Bruyninckx, H.; De Schutter, J. Nonlinear Kalman Filtering for Force-Controlled Robot Tasks 280 p. 2005 [3-540-28023-5] Vol. 18: Barbagli, F.; Prattichizzo, D.; Salisbury, K. (Eds.) Multi-point Interaction with Real and Virtual Objects 281 p. 2005 [3-540-26036-6]

Vol. 10: Siciliano, B.; De Luca, A.; Melchiorri, C.; Casalino, G. (Eds.) Advances in Control of Articulated and Mobile Robots 259 p. 2004 [3-540-20783-X] Vol. 9: Yamane, K. Simulating and Generating Motions of Human Figures 176 p. 2004 [3-540-20317-6] Vol. 8: Baeten, J.; De Schutter, J. Integrated Visual Servoing and Force Control 198 p. 2004 [3-540-40475-9] Vol. 7: Boissonnat, J.-D.; Burdick, J.; Goldberg, K.; Hutchinson, S. (Eds.) Algorithmic Foundations of Robotics V 577 p. 2004 [3-540-40476-7]

Vol. 17: Erdmann, M.; Hsu, D.; Overmars, M.; van der Stappen, F.A (Eds.) Algorithmic Foundations of Robotics VI 472 p. 2005 [3-540-25728-4]

Vol. 6: Jarvis, R.A.; Zelinsky, A. (Eds.) Robotics Research { The Tenth International Symposium 580 p. 2003 [3-540-00550-1]

Vol. 16: Cuesta, F.; Ollero, A. Intelligent Mobile Robot Navigation 224 p. 2005 [3-540-23956-1]

Vol. 5: Siciliano, B.; Dario, P. (Eds.) Experimental Robotics VIII 685 p. 2003 [3-540-00305-3]

Vol. 15: Dario, P.; Chatila R. (Eds.) Robotics Research { The Eleventh International Symposium 595 p. 2005 [3-540-23214-1]

Vol. 4: Bicchi, A.; Christensen, H.I.; Prattichizzo, D. (Eds.) Control Problems in Robotics 296 p. 2003 [3-540-00251-0]

Vol. 14: Prassler, E.; Lawitzky, G.; Stopp, A.; Grunwald, G.; Hagele, M.; Dillmann, R.; Iossiˇdis. I. (Eds.) Advances in Human-Robot Interaction 414 p. 2005 [3-540-23211-7]

Vol. 3: Natale, C. Interaction Control of Robot Manipulators { Six-degrees-of-freedom Tasks 120 p. 2003 [3-540-00159-X]

Vol. 13: Chung, W. Nonholonomic Manipulators 115 p. 2004 [3-540-22108-5] Vol. 12: Iagnemma K.; Dubowsky, S. Mobile Robots in Rough Terrain { Estimation, Motion Planning, and Control with Application to Planetary Rovers 123 p. 2004 [3-540-21968-4] Vol. 11: Kim, J.-H.; Kim, D.-H.; Kim, Y.-J.; Seow, K.-T. Soccer Robotics 353 p. 2004 [3-540-21859-9]

Vol. 2: Antonelli, G. Underwater Robots { Motion and Force Control of Vehicle-Manipulator Systems 209 p. 2003 [3-540-00054-2] Vol. 1: Caccavale, F.; Villani, L. (Eds.) Fault Diagnosis and Fault Tolerance for Mechatronic Systems { Recent Advances 191 p. 2002 [3-540-44159-X]

Henrik I. Christensen (Ed.)

European Robotics Symposium 2006 With 89 Figures

Professor Bruno Siciliano, Dipartimento di Informatica e Sistemistica, Universit`a degli Studi di Napoli Federico II, Via Claudio 21, 80125 Napoli, Italy, email: [email protected] Professor Oussama Khatib, Robotics Laboratory, Department of Computer Science, Stanford University, Stanford, CA 94305-9010, USA, email: [email protected] Professor Frans Groen, Department of Computer Science, Universiteit van Amsterdam, Kruislaan 403, 1098 SJ Amsterdam, The Netherlands, email: [email protected]

Editor Professor Dr. Henrik I. Christensen Center for Autonomous Systems CAS/CVAP/NADA Kungliga Tekniska Hoegskolan (KTH) 10044 Stockholm Sweden

ISSN print edition: 1610-7438 ISSN electronic edition: 1610-742X ISBN-10 3-540-32688-X Springer Berlin Heidelberg New York ISBN-13 978-3-540-32688-5 Springer Berlin Heidelberg New York Library of Congress Control Number: 2006921135 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, speciﬁcally the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microﬁlm or in other ways, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under German Copyright Law. Springer is a part of Springer Science+Business Media springeronline.com © Springer-Verlag Berlin Heidelberg 2006 Printed in Germany The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a speciﬁc statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: Digital data supplied by editor. Data-conversion and production: PTP-Berlin Protago-TEX-Production GmbH, Germany Cover-Design: design & production GmbH, Heidelberg Printed on acid-free paper 89/3141/Yu - 5 4 3 2 1 0

Editorial Advisory Board EUROPE Herman Bruyninckx, KU Leuven, Belgium Raja Chatila, LAAS, France Henrik Christensen, KTH, Sweden Paolo Dario, Scuola Superiore Sant’Anna Pisa, Italy R¨udiger Dillmann, Universit¨at Karlsruhe, Germany AMERICA Ken Goldberg, UC Berkeley, USA John Hollerbach, University of Utah, USA Lydia Kavraki, Rice University, USA Tim Salcudean, University of British Columbia, Canada Sebastian Thrun, Stanford University, USA

EUR ON

ASIA/OCEANIA Peter Corke, CSIRO, Australia Makoto Kaneko, Hiroshima University, Japan Sukhan Lee, Sungkyunkwan University, Korea Yangsheng Xu, Chinese University of Hong Kong, PRC Shin’ichi Yuta, Tsukuba University, Japan

European

***

***

STAR (Springer Tracts in Advanced Robotics) has been promoted under the auspices of EURON (European Robotics Research Network)

Research Network

***

***

ROBOTICS

An Ontology of Robotics Science John Hallam1 and Herman Bruyninckx2 1 2

Mærsk Institute, University of Southern Denmark, Campusvej 55, DK–5230 Odense M, Denmark [email protected] Dept Mechanical Engineering, Katholieke Universiteit Leuven, Celestijnenlaan 300B, B-3001 Leuven, Belgium. [email protected]

Summary. This paper describes ground-breaking work on the creation of an ontology for the domain of robotics as a science. An ontology is a collection of terms, concepts and their inter-relationships, represented in a machine-usable form. An ontology in a particular domain is useful if the structure it adds to the domain is simple enough to be understood quickly and intuitively, and rich enough to increase insight into the whole domain to a level where this increased insight can lead to innovation and increased eﬃciency in scientiﬁc and practical developments. This paper presents an ontology for the science of robotics, and not for robots as objects: the latter ontology describes the physical and technical semantics and properties of individual robots and robot components, while the ontology of the science of robotics encodes the semantics of the meta-level concepts and domains of robotics. For example, surgical robotics and industrial automation are two concepts in the ontology of the science of robotics, while the semantics of robot kinematics and dynamics, or of a particular robot control algorithm belong to the ontology of robots as objects. The structure in the presented ontology for the science of robotics consists of two complementary sub-structures: (i) the robot agent and robot system models (i.e., what components are required in a robot device, and in a robotic application, respectively), and (ii) the Context Space (i.e., ordinal and categorical relationships between physical and computational aspects of robot agents and systems in which the sub-domains of robotics can be mapped out). The implementation of the Context Space concept using standard ontology tools is explained. The paper illustrates its expected usefulness with examples of sub-domains of robotics expressed as contexts in the context space, and with two use cases for the ontology: (i) the classiﬁcation of conference or journal paper submissions, and (ii) the guidance of new researchers into the domain of robotics.

1 Introduction 1.1 Robotics as a Science Robotics is to a large extent a science of integration, constructing (models of) robotic systems using concepts, algorithms and components borrowed from

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 1–14, 2006. © Springer-Verlag Berlin Heidelberg 2006

2

J. Hallam and H. Bruyninckx

various more fundamental sciences, such as physics and mathematics, control theory, artiﬁcial intelligence, mechanism design, sensor and actuator technology. The function and properties of a robotic system depend on the components from which it is made — the speciﬁc sensors, actuators, algorithms, mechanism — but, beyond that, they depend on the way those components are integrated. Furthermore, a full description of a robotic system must include information about the task it is to perform and the environment in which the task is to be undertaken. As we shall see, this view of robotics (i.e., the system formed by the robot agent, its task and the environment in which it has to perform that task) is crucial in developing a full ontology for robotics. 1.2 Ontologies for Robotics An ontology is a formal deﬁnition of concepts, terms and relationships appropriate to some domain of knowledge, generally expressed in a formalism that allows machine-usability of the encoded knowledge. (The Wikipedia contains a good introduction to the concept of an ontology, [3].) Ontologies typically serve two purposes: 1. providing agreed and unambiguous terminology for a domain, with the goal of helping humans express, transform and transfer their knowledge more eﬀectively and accurately. 2. allowing automatic use of that knowledge, for instance in exchange or linking of data between processes, or translation of terms between languages. The latter purpose is sometimes better known under the name Semantic Web, [1], and the W3C have deﬁned standards for computer-readable representation of ontologies, [2]. Robotics, like other sciences, has need of suitable ontologies. For instance, an agreed ontology for sensor data would make possible much greater interoperability between sensing and other modules in a robotic system. Individual modules could label the data they generate using terms from the agreed ontology; other processes would then have available a deﬁnite semantics for the data. Similar applications can be envisaged in the domains of mechanical/electronic and control/software engineering in robotics. Such ontological knowledge would also be of use for designers of robotic systems when selecting and matching components. We describe this kind of ontology as an ontology of robots as objects. While a complete ontology of robotics requires such knowledge, the source of that knowledge is largely in the sciences and technologies on which robotics builds, viz. physics, control engineering, and so on. Such an ontology may be extremely valuable, but is not the focus or purpose of this paper. However, the ontology presented in this paper could be connected to these object ontologies without problem.

An Ontology of Robotics Science

3

Robotics, in its guise as a science of integration, implies a second kind of ontology. Consider a typical robotics conference, at which one may ﬁnd sessions or talks on “ﬁeld robotics” or (perhaps) “factory automation”. Roboticists use such labels as a taxonomy of their ﬁeld, and have a more-or-less clear, but informal, understanding of the signiﬁcant problems, opportunities and special requirements of each sub-ﬁeld: of the context for the particular sub-domain. So, in the context of papers presented at conferences, even the short labels “ﬁeld robotics” or “factory automation” are suﬃcient to help reviewers and readers focus their attention to the contributions of the paper in the topics that are supposed to be diﬃcult and relevant in the state of the art in those domains. It takes students and practitioners in robotics quite some time to learn the links between the short labels on the one hand, and the scientiﬁc challenges and contributions they can expect in papers labelled in that way. This paper makes a ﬁrst contribution in making these links more explicit and structured, in order to not only speed up the learning phase of human practitioners but also to facilitate (semi) automatic support for selection of reviewers, categorization of papers or research proposals, web or library searches, etc. The starting point of the presented research is thus that there is knowledge about the structure of the robotics science itself — what sub-domains are there; how are they inter-related; how can we express the relationships between sub-domains in a principled way? Once again, the idea of an ontology can help. We call an ontology dealing with this knowledge an ontology of robotics science. The existence and formalisation of that ontology is the key insight and focus of the present paper. 1.3 Organisation of the Paper In the following sections we motivate, illustrate and explore the idea of an ontology of robotics science. Section 2 considers robotic systems in general: comprising the agent, the task and the environment. A model of the components expected in a robotic system is presented. Section 3 then presents the key concepts in the ontology of robotics science, particularly the notion of scaling laws, which structure the domain of robotics. Section 4 investigates how the proposed ontology might be applied, and demonstrates its potential usefulness. Section 5 illustrates how the key concept of Context Space can be implemented using standard ontology tools. The ﬁnal section summarises the paper’s conclusions and indicates planned future work.

2 A Model of Robotics Representing human knowledge about technological domains, such as robotics, is equivalent to deﬁning sets of models with which humans can analyse the real world, make calculations about it, make predictions, etc. In other words,

4

J. Hallam and H. Bruyninckx

a model is a mathematical simpliﬁcation of the real world, representing only those aspects of the real world that are relevant to a particular human purpose. The purpose of this paper is to bring structure to human knowledge of robotics, but not to give detailed technological descriptions of the workings of robotics devices or algorithms. We believe that the graphical models in ﬁgure 1 and 2 provide us with the appropriate balance between detail and generality for the purpose of this paper. Figure 1 gives a “robot-centric” model of robotics: it depicts a robot “agent” as an integration of mechanical hardware, sensing hardware and actuation hardware, with planning software, control software, sensor processing software and actuator drive software. A successful robot interconnects knowledge and hardware at “appropriate” scales and with “appropriate” interfaces. Later sections of the paper will indicate more clearly what “appropriate” really means, and make clear that terms such as “ﬁeld robotics” or “surgical robotics” indicate particular sub-manifolds of the “robot-centric” continuum (and the “application-centric” continuum of ﬁgure 2, see below) that have proven suﬃciently successful or interesting to attract the intense research efforts from large communities.

Fig. 1. The “robot-centric,” or agent-model, of robotics.

Fig. 2. Robotic systems as the manifold of robot agents, tasks and environments.

Figure 2 gives an “application-centric” or “systems-centric” view of robotics complementary to the previous ﬁgure: it makes clear that success of a robot agent is not a property of the agent in itself, but is determined also by (i) the task the robot is expected to perform, and (ii) the environment in which that task is to be performed. Each successful robotic systems domain covers a certain subspace of the triangular continuum depicted in ﬁgure 2. In some

An Ontology of Robotics Science

5

cases, it is the robot device that contributes most to the success of the system, and its subspace is situated close to the agent vertex of the triangle. In other systems, it is the environment (“hardware”) that is most important, for example in sub-sea or airborne robots where water or air provide the necessary support to the robot. In yet other systems, the task is the determining factor.

3 An Ontology of Robotics Science The key to deﬁning relationships and structure between the various subﬁelds of robotics is an understanding of how the subﬁelds come about. Recall that a particular subﬁeld expresses similarities between a class of robotic systems, and implies relevant background knowledge, signiﬁcant problems, useful properties and so on. What accounts for much of this variation is scaling laws. Consider the diﬀerence in the ﬂight of an air-plane and a bumblebee. Each exploits diﬀerent aerodynamic properties accessible to it because of the physical scale (tens of metres vs. millimetres) in which it operates: air behaves quite diﬀerently in diﬀerent physical circumstances. This diﬀerence is an expression of a physical scaling phenomenon wherein physical system properties depend on the characteristic length of the system. Analogues of this principle can be constructed in other physical scales, such as time or mass, and in more abstract scales such as computational complexity. It is the placement with respect to a set of such scales that determines the properties, problems and possibilities of a given subclass of robotic system. 3.1 Context Space The Context Space is a conceptual space spanned by the various scaling dimensions available to a robotic system (or to its designer). Since, as we have argued in section 2, such a system comprises a device or agent, an environment, and a task, each of these essential components contributes scale dimensions and constraints to the system. In this section we examine some of these dimensions. The example dimensions we have chosen are intended to be illustrative rather than complete — space precludes an exhaustive listing of all possibilities, and such is not necessary to make our point. Each essential component of a complete robotic system contributes scaling dimensions. Thus device scales are the most straightforward to specify. The characteristic length and time constant of the robotic system, its mass, the number of components it comprises and the extent to which they interact with each other are important characteristics, as are the number of controllable degrees of freedom and the total degrees of freedom of the system.

6

J. Hallam and H. Bruyninckx

environment scales deﬁne the type and properties of the environment in which the task is carried out. There are two classes of environment: bulk and interface. Examples of the former are vacuum (as in space robotics) or liquid (as in underwater robotics). Bulk environments can be ordered by their mechanical impedance. Examples of the latter, interface, type include solid-gas (e.g. a table top, or the outdoor land surface) or liquid-gas (e.g. the surface of the sea). Salient properties here include the size and spatial frequency of surface variation — how smooth the interface is. The size of the environment (its characteristic length); the speed with which events occur (its characteristic time constant); and whether the environment is abiotic or biotic (or on the interface of the two) constitute other examples of relevant dimensions of variability. task scales include the spatiotemporal properties of the task, for instance its characteristic precision and the lifetime of the robotic system when working on the task, and the risk cost, i.e. the cost associated with failure or error in the task. The latter can informally be deﬁned as the cost of ﬁxing a disaster caused by system failure. For laboratory robotics it is, relatively speaking, very small; for surgical or space robotics it is relatively large. Many of these scales have an ordinal structure, that is, the axis representing the dimension has a ‘small value’ end and a ‘large value’ end, with intermediate values that may be arranged in order of size. As we shall see below in section 3.3, these ordering relationships along individual axes contribute to the rich structure of the context space. 3.2 Contexts Contexts are the basic mechanism for using the ontology. They provide a means to distinguish groups of robotic systems — and so diﬀerent subﬁelds of robotics as a whole — whose properties are restricted to sub-ranges of certain scale axes. For example, nanorobots have a device length scale in the range 10−9 –10−7 metres. More generally, we can specify sets of systems whose properties lie within speciﬁc ranges on several axes simultaneously; most generally, a context may be a union of such convex regions in the space. Conceptually, a context is a sub-class of the class of all possible robotic systems. Such classes will usually (but not necessarily) have intuitive names, for example “ﬁeld robotics,” “medical robots,” “swarm robots,” and each context implies a set of relevant technology, background knowledge, scientiﬁc problems and solutions following from the particular choices of scale which deﬁne the context. 3.3 Relationships Between Contexts Contexts, as deﬁned above, can be thought of as sub-classes of the class of robotic systems; sub-classes generated by restricting the values of certain scaling properties (e.g. device length, risk cost, number of components, and so

An Ontology of Robotics Science

7

on) to particular values or sets of values from their respective scales (e.g. millimetre-metre, EUR 100–10000, 5–10, etc.). As such, contexts can take part in the standard class relationships of set theory: they can be super- or sub-classes of other classes; intersections, unions and complements can be deﬁned; and membership can be checked. However, more interesting are the relationships implied by the ordinal structure of the scale axes. For instance, robot devices whose physical size (characteristic length) is in the centimetre-decimetre range are neighbours of those in the decimetre-metre range: they will have many properties in common. On the other hand, nanometre scale robotic devices will be quite different from centimetre ones, in some respects, since the dominant physical eﬀects diﬀer between those characteristic sizes. Neighbouring contexts are often contexts between which there exist loose analogies or trade-oﬀs. For instance, one could describe a 100 mobile robot collective as a group of 100 6-dof systems or a single 600-dof system: exchange physical dimensions for interconnected system piece complexity. Neighbourhood implies quasi-invariance between aspects of the relationships that describe the same system in diﬀerent contexts. Change in one scaling dimension is compensated by a corresponding change in another dimension. Suﬃciently sophisticated reasoning mechanisms could exploit neighbourhood relationships to propose novel connections between robotic systems, or identify possible similarities between hithertofore disjoint subﬁelds of the discipline. The neighbourhood relationships provided by the context scales also force human students in robotics to think more explicitly about the really unique, respectively shared, properties of speciﬁc robotics domains.

4 Using the Ontology The proposed ontology is unusual in that it focusses on the structure of a ﬁeld of science rather than on the objects with which that ﬁeld is concerned. Nevertheless, such an ontology has considerable practical use. In addition, it is easier to start with the ontology of robotics as a science, because, as presented in this paper, this ontology can be explained and understood (almost) completely in a relative short amount of time. In this section we give three illustrative examples of practical usefulness of the ontology. 4.1 Structuring the Domain of Robotics Science To illustrate the ideas presented so far, consider the various sub-ﬁelds of robotics typically found in the contents listings of conference proceedings: can we deﬁne them using context space ideas and axes? Suppose we start with Field Robotics. We can say that the device scale (robot size) is in the metre range; the environment length scale is 102 to 1010

8

J. Hallam and H. Bruyninckx

metres; the device/mission lifetime ranges between 103 and 1010 seconds; the environment may be vacuum, gas, liquid or an interface. So far so good; now consider underwater robotics. This is Field Robotics in a bulk liquid environment. Space Robotics is Field Robotics in a bulk vacuum environment, with environment length scale from 105 to 1010 metres, device/mission lifetime 106 –1010 seconds, say (for space craft) and a high risk cost. Consider Research Lab Robotics. In this case we have device lengths of 10−3 –100 metres; environment length scale 101 metres, say; gas or smooth gas-solid interface; lifetime in the 102 –105 second range; and low risk cost. Table-top robotics, for instance using the popular Khepera robot, is a specialisation of the Research lab context, with 10−2 metre device scale, 100 metre environment scale, very smooth solid-gas interface environment, very low risk factor, low environmental interaction and lifetime in the 102 –103 second range. As a ﬁnal example, consider factory automation or industrial robotics. Here we have device scales in the 10−3 –100 range, environments in the 100 –101 metre range, gas-solid interface, lifetimes from 105 –107 seconds; medium-high risk cost; moderate system complexity and low environment interaction. Although these examples are still somewhat vague — recall, for instance, ‘medium-high risk cost’ or ‘moderate system complexity’ — they nevertheless illustrate that the diﬀerent sub-ﬁelds of robotics occupy speciﬁable parts of the context space and that related sub-ﬁelds occupy neighbouring or subset regions of the context space. Thus it is possible to deﬁne the subﬁelds of robotics by labelling the appropriate regions of the context space and thereby infer relationships between subﬁelds that might not immediately be apparent from just their names. In addition, being confronted with the diﬀerent ordinal scales that are relevant in a particular robotics domain will stimulate students and researchers to think about how the fundamental physical properties of a particular context could change when “moving” to a neighbouring context, or they could get inspiration from successful solutions implemented in neighbouring domains that they had not thought of before. 4.2 Use Case: Guided Reading for New Roboticists All teachers in robotics have experienced the following evolution in their students: the ﬁrst couple of weeks or months in their study of robotics, students are reading tons of papers, working hard to see the forest for the trees, but are mostly unable to evaluate the essential contributions (or lack thereof) of every individual paper they read. Good teachers also know that they must organise regular reading sessions with their students, in order to guide this process of getting acquainted with a particular robotics research domain. We believe that a key role of the teachers in this coaching phase is to indicate to their students what are the relevant relationships between the details of the science explained in particular papers, and the goals of making

An Ontology of Robotics Science

9

a better robot system (i.e., the combination of agent, task and environment) in the particular domain of interest to the student. We believe that teachers (most often) implicitly convey to their students the kind of ontology that this paper presents, and that making these ontological relationships explicit will help the students in their learning. In addition, by using the ontology of this paper in real-world reading classes, the ontology will become more complete and reﬁned. In addition, it could possibly become the basis for a series of textbooks that cover the whole domain of robotics in the most eﬃcient way. 4.3 Use Case: Classiﬁcation of Papers and Reviewers A second application of these ideas is to paper classiﬁcation. Imagine that authors submitting papers to a robotics journal or conference (or, indeed, a funding agency) are invited to indicate approximately where, on ﬁgure 2 and ﬁgure 1, their paper content falls. Then, depending on their answer, they are invited to specify where it lies on various appropriate scale axes. For example, the author of a paper on inertial guidance for underwater robots would indicate ‘agent-centred’, ‘sensors’, and choose ‘bulk liquid environment’, ‘environment size 102 –103 metre’, ‘lifetime 103 –104 second’, and so on. Reviewers could similarly indicate their areas of expertise. Software based on the ontology could then match reviewers to papers depending on their distance in the contextual space. For most papers, this would produce the expected results: a paper on space robotics, for instance, would go to space robotics referees. But interesting desirable eﬀects are possible. A paper on collective robotics might be oﬀered to a reviewer interested in the control of very highly redundant systems, on the arguably sensible grounds that 100 6-dof robots have some similarity with one 600-dof one — the two subjects are neighbours in (some dimensions of) the context space.

5 Implementing the Ontology of the Science of Robotics To be useful, it must be possible to implement the ontology of the science of robotics using the standard ontology tools of the semantic web. The World Wide Web Consortium has selected OWL as its standard for the representation of ontological knowledge [4]. OWL is built on top of standard infrastructure (RDF and XML) and provides mechanisms for representing concepts (classes), instances, properties and relationships. It comes in three ﬂavours — OWL-Lite, OWL-DL and Full OWL — which represent diﬀerent compromises between representational power and computational tractability. The ontology described here is implemented using OWL-DL.

10

J. Hallam and H. Bruyninckx

5.1 Implementing the Context Space The key concept to implement is the Context Space. To do that, we identify a context with the set of all robotic systems falling into that context; i.e. ‘Space Robotics,’ in terms of the ontology, is deﬁned as the class of robotic systems studied by that ﬁeld. The class of such systems is actually deﬁned using more fundamental properties, such as the environment, spatial and temporal scales in which the systems operate, and so on. The basis for implementing the Context Space is then to deﬁne the class of all robotic systems, sub-classes of which represent individual contexts. Contexts are deﬁned using standard class operations of union, intersection, complement and property value restriction (restricting the legal values of a speciﬁc instance property to a known class). In line with the discussion of section 2, we therefore assert that each robotic system (an instance of the robotic System Class) comprises a device (the agent), an environment and a task. We further deﬁne three classes, which we call Aspect Classes, consisting respectively of all devices, all environments and all tasks, as the value ranges of the three instance properties of a system. Associated with the instances of each aspect class are scaling properties, for example device-characteristic-length or environment-impedance, the values for which are drawn from Scale Classes representing the various scaling axes. Notice that diﬀerent aspects may have properties relating to the same scale: device, environment and task characteristic lengths or times would be an obvious example. Figure 3 illustrates this general scheme. Technically, each instance of the robotic Systems class has three properties, whose values are restricted to lie each in one of the three aspect classes; and each instance of an aspect class (for example, a particular environment) has scaling properties whose values are restricted to particular sub-classes from given scale axes. This construction has the crucial advantage that the class structure is deﬁned intensionally — we are not required to enumerate all robotic systems, all environments, etc. — while also allowing speciﬁc systems, devices and so on to be represented as instances. To illustrate the implementation strategy, consider systems using nanoscale robotic devices, which can be characterised by the fact that their devicelength scale is in the nanometre range. Figure 4 illustrates how this is represented: a sub-class of the length scale class is deﬁned to represent the nanometre range, and a sub-class of the devices aspect class is deﬁned such that the device-length property is restricted to take values from the nanometre range class. Finally, a sub-class of the systems class with device property values restricted to the just-deﬁned nano-device class completes the representation of nano-scale robotic systems. Notice that we have not restricted the environments in which these systems may operate nor the tasks to which they may be applied. The red (dashed) components in the ﬁgure illustrate the newly

An Ontology of Robotics Science

11

Fig. 3. Class hierarchy in the Robotics-as-Science Ontology: the universe is represented as the class of all robotic Systems, where each system (instance) has three Aspects — device, environment and task. Instances of the Aspect Classes have scaling properties, e.g. characteristic length, with values from the Scale Classes.

Fig. 4. Nano-Robotic Systems

Fig. 5. Medical Nano-Robotics

deﬁned sub-classes; the standard OWL reasoners are able to infer that these are indeed sub-classes of their parents. Figure 5 takes the example one stage further. Consider medical nanorobotics. The principal distinction between this and nano-systems robotics, for our illustrative purposes, is the environment in which the systems must work — on a macro-molecular scale and in interaction with living matter. This is represented by the construction of a sub-class of biotic nanoscale environments to which the environment property of the system is restricted. The red (dashed) components in the ﬁgure illustrate these new deﬁnitions. Once again, standard OWL reasoners can infer that the medical nano-systems class

12

J. Hallam and H. Bruyninckx

is a sub-class of the nano-robotic systems or, equivalently, that medical nanorobotics is a sub-ﬁeld of nano-scale robotics. 5.2 Context Relationships The ability to represent contexts is crucial to the ontology, but so is the ability to represent or infer relationships. As we have seen, certain relationships such as class inclusion can be inferred by the standard OWL reasoners. Not so with neighbourhood relationships, however. The problem is that OWL reasoners contain no machinery for inferring ‘distance’ between classes. To illustrate this, consider the length scale. We can construct in OWL a partition of this scale into (amongst others) millimetre, centimetre, decimetre and metre lengths with the appropriate ordering relationships (e.g., centimetres are longer than millimetres and shorter than metres). But it is not possible with standard reasoners to infer that the metre sub-class is an indirect neighbour of the centimetre class two steps up the ordinal scale. What this means is that neighbourhood relationships, which depend on the ‘distances’ between context properties in the ordinal scale axes, cannot be inferred by standard reasoners since the ordinal ‘distances’ themselves cannot be inferred by those reasoners. However, more specialist reasoners can be built to make such inferences, and one can imagine such reasoners working on the ontology enumerating neighbourhood relationships between the represented contexts and automatically adding to the ontology explicit representations of discovered neighbourhood relationships. 5.3 Summary The ontology of the science of robotics depends on the key concept of contexts, which we can straightforwardly and intuitively represent using the standard ontology language OWL by deﬁning a context (a sub-ﬁeld of robotics) to be represented by the class of all systems belonging to that context (sub-ﬁeld). Standard reasoners can determine simple class relationships between contexts. Neighbourhood relationships between contexts cannot be inferred using standard reasoners, but specialist discovery processes could be implemented to annotate the ontology with such relationships.

6 Conclusions and Future Work 6.1 Conclusions The present paper has motivated the notion of an ontology of the science of robotics as opposed to an ontology for the objects of which robotic systems are composed. An ontology of robotics science allows us to deﬁne terms and concepts such as “surgical robotics,” “ﬁeld robotics,” or “nanorobotics” in an

An Ontology of Robotics Science

13

objective way in terms of more primitive technological concepts. This allows us to give formal (machine-usable) meaning to these various terms and to explore the relationships between them. These meanings and relationships provide a meta-level structure in the scientiﬁc discipline of robotics, that could be exploited in various ways. The formal setting in which such deﬁnitions are made is a Context Space established by considering the physical, environmental and task-related scaling laws and relationships that apply to robotic systems. New sub-ﬁelds of robotics, and new relationships between them and between existing subﬁelds, may be discovered by considering the relationships between them in context space. The ontology has been implemented using the standard ontology language OWL-DL using the Proteg´e suite. The proposed ontology is believed to be simple enough to be understood by humans within the span of a couple hours, while at the same time it is rich enough to bring non-trivial and hence useful structure to scientiﬁc domain of robotics. The ontology may be used, inter alia, as a tool for classiﬁcation of robotic material, the matching of reviewers to conference, journal or funding agency submissions, or during teaching robotics to new students. Although the description of the ontology in the paper is somewhat incomplete, this is natural: to our knowledge there is no state of the art at all addressing this kind of ontology for robotics and also no truly comparable work in other (technological or non-technological) domains. Hence the goal of the paper is to present in outline a novel but, we believe, valuable tool for roboticists. 6.2 Future Work Before an ontology of robotics science is complete and fully usable, an amount of future work remains — both conceptual and practical. Conceptual work includes elaborating the set of scale axes to identify a complete, minimal and consistent set; identiﬁcation of the kinds of questions and use cases for which the ontology provides a useful resource; and wide consultation with the robotics community on the clarity, utility and completeness of the conceptual framework. Practical work comprises completing the implementation of the concepts and relationships of the ontology tools conforming to the standards and recommendations established by the World-Wide Web Consortium; and the formal deﬁnition and implementation of suitable query and reasoning interfaces to the implemented ontology. The ontology must also be tested more intensely on the “educational workﬂoor” of teaching robotics to masters and PhD level students.

14

J. Hallam and H. Bruyninckx

Acknowledgement The authors gratefully acknowledge the support of the EURON Research Atelier Rose: a robotics ontology for the semantic web, and the many stimulating discussions with members of EURON’s Education and Training committee, which have, to a large extent, helped shape the ideas of the ontology presented in this paper. Wei Xu contributed many useful ideas to the described implementation of the ontology.

References 1. Grigoris, A. and Van Harmelen, F. A Semantic Web Primer, MIT Press, 2004. 2. W3C. Semantic Web, http://www.w3.org/2001/sw/, accessed on-line on September 30th, 2005. 3. Wikipedia. Ontology (computer science), http://en.wikipedia.org/wiki/ Ontology (computer science), accessed on-line on September 30th, 2005. 4. W3C. OWL Web Ontology Language Overview, http://www.w3.org/TR/ owl-features/, accessed on-line on January 12th, 2006.

A Multi-agent System Architecture for Modular Robotic Mobility Aids Georgios Lidoris and Martin Buss Institute of Automatic Control Engineering Technische Universit¨ at M¨ unchen D-80290 Munich, Germany [email protected]

Abstract In this paper a multi-agent system architecture for a modular mobility enhancement system is presented. The system consists of one or multiple mobile robotic platforms and a set of user deﬁned application modules, such as chairs, tables multifunctional chairs etc. All these modules can be inexpensive everyday life items, that become functional when a mobile platform is properly attached to them. This way the system can act as a wheelchair, as a carrier of objects, or even as a walker. ∗

1 Introduction With the increase in the number of senior citizens, there is a growing demand for mobility aids. Older people have problems moving themselves, handling or moving objects and at the same time operating complicated devices. Therefore mobility aids with intelligent capabilities become a necessity. Several robotic, intelligent wheelchairs have been proposed demonstrating navigation, manipulation and transportation capabilities. Examples include Rolland [1] , TAO [2], OMNI [3], NavChair [4] and many other. However existing systems are still diﬃcult to operate, to adjust to the needs of the individual user and are very costly. At the same time research on multi-agent systems has provided both principles for the construction of complex systems involving multiple agents and mechanisms for the coordination of independent agent behaviours. In this paper we intend to apply well-analysed concepts from multi-agent and multi-robot systems research, to create a system architecture for a modular mobility enhancement aid. This is a new and very demanding application ∗

This work was partially supported by the Speciﬁc Targeted Research Project MOVEMENT co-funded by INFSO DG of the European Commission (contract number 511670)

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 15–26, 2006. © Springer-Verlag Berlin Heidelberg 2006

16

G. Lidoris and M. Buss

domain for multi-agent systems and besides presenting the proposed approach we intend to expose its speciﬁc requirements. The system can consist of one or multiple mobile robotic platforms and a set of user deﬁned application modules, such as chairs, tables, multifunctional chairs etc. All these modules can be inexpensive everyday life items, that become functional when a robotic mobile platform is properly attached to them. This way the system can act as a wheelchair, as a carrier of objects, or even as a walker. Also multiple levels of autonomy are supported by the system architecture, from manual operation to autonomous functionality. Such a modular system, addresses many diverse needs and can be used by people with diﬀerent mobility problems. Its scalable multi-agent architecture allows it to be deployed for the home care of an individual, as well as in large health care institutions.

2 Related Work Several approaches have been proposed in the multi-agent community to coordinate agent actions. Distributed behaviour-based architectures have been developed, like ALLIANCE [5] and [6], which were capable of dealing with complex tasks but lacked on scalability and robustness. Other approaches address multi-robot cooperation through negotiation, such as M+[7], Traderbots [8], MURDOCH [9] and numerous others. Market-based architectures are winning popularity and already a large amount of literature relevant to the ﬁeld exists. An interesting survey of market based multi-robot coordination can be found in [10]. Our approach uses a negotiation mechanism for task allocation, as well as a behavioural-based layer for the execution of complex tasks. This way the robustness of market-based mechanisms is combined with the ability of behaviour-based approaches to deal with tasks of increased complexity. A contribution to the challenges of multi-agent systems research as described in [10] is made, by introducing a new real-world application domain where long-term, reliable and robust operation of human, robot teams can be demonstrated.

3 Concept of System Architecture As we mentioned above our application domain introduces several special requirements. We are dealing with a safety critical domain, where we have to ensure the continuous functionality of our system. Direct task allocation and supervision of system’s decisions by the user must be supported. Tasks arrive in an unpredicted manner, from several user interfaces. Finally new modules can be inserted to and removed from the system at any time.

A Multi-agent System Architecture for Modular Robotic Mobility Aids

17

The system architecture is responsible for the coordination of modules consisting the system. It shall enable the user dependent usage of the modules and has the task of controlling the interconnection of modules and their behaviour. More speciﬁcally we present a Multi-Agent System architecture which is responsible for receiving high-level commands from the user interface, decomposes them and delegates sub-tasks to the appropriate application modules. These modules are dynamically selected according to utility functions, which are created using performance indicators and task requirements. In the formal multi-robot task allocation (MRTA) framework introduced in [11], our system addresses ”Single-task robots, single robot tasks, instantaneous assignment” (ST-SR-IA). Furthermore our system continuously runs background tasks such as managing module power levels, performing conﬂict resolution between system modules during task executions and ﬁnally enabling the system’s autonomous functionalities by making use of a behaviour-based task model. Component selection for task allocation is performed by using a simple distibuted ﬁrst-price one round auction protocol, similar to the Contract-Net protocol [12]. The progress of task execution is monitored by the winner and if no progress is made, then the task is re-assigned through a new negotiation. This is a very important fault-prevention functionality of the system. 3.1 Agent Model Taking into account the distributed nature and complexity of the application domain, we decided to split the control of our system amongst a number of deliberative agents. Each agent is capable of reasoning, according to the classical hybrid architecture paradigm [13]. At the same time, by distributing control to a number of agents, we achieve robustness and scalability gains. Therefore our system can be considered as an extension of the layered approach [14], [15], [16] to a multi-agent domain, exploiting advantages of both centralised and distributed approaches. The general internal structure of a system agent is illustrated in Figure 1. It consists of three parts. The ﬁrst one is responsible for communicating with other agents and handling communication requests. The second part is a scheduler, which receives task requests, creates a queue and schedules their execution from the third layer. The task execution layer is responsible for the agent’s actions, which according to each agent’s role can vary from simple communication acts to movements, for our physical agents. By using such an internal organization, we are able to deal with coordination and cooperation issues that arise between our agents. Our system becomes capable of having physical agents that respond quickly to dynamic events e.g. collisions, while at the same time producing and executing complex strategies for achieving multiple goals. This is due to the fact that each system component is able to schedule its own actions according to its internal capabilities and state.

18

G. Lidoris and M. Buss communication

scheduling

task execution

Fig. 1. Internal structure of a system agent

3.2 Overview of the Proposed Approach The proposed system architecture consists of several software agents that are responsible for performing task decomposition, task allocation through negotiations as well as system monitoring and module management. It also contains physical agents which represent the modules that are available to the system at each time point. No limitations exist to the number of possible modules that can be inserted to the system, enabling it to be used in small setups with only one mobile base and a few application modules, as well as in large care institutions with an arbitrary number of mobile bases and application modules. In Figure 2 we present an overview of our architectural concept, for a setup where the available modules that can be docked with the platforms are chairs, tables, multifunctional chairs and walker modules. When the user issues a command via the input hardware of the User Interface, this command is integrated and preprocessed and then sent to a link layer. This link layer between the Multi-Agent System and the User Interface is responsible for transforming the issued command, into an appropriate form in order to be transmitted to the control architecture. Important information being encapsuled are: • • • •

Level of autonomy in which the system is currently operating Type of operation to be performed by the system Which application module issues the command Command speciﬁc parameters e.g. goal position coordinates, in the autonomous functionality or speeds to be executed by the mobile platform, read by the input device for the manual functionality

This high-level command is transmitted from the user interface link layer, to the system. In order to make clear how a command is handled by the system, we have to describe each agent’s role in more detail.

A Multi-agent System Architecture for Modular Robotic Mobility Aids

19

User Interface Link Layer High-level commands / notifications

System State Agent

Mediator Agent

Task Allocation Agent

Platforms

Chairs

Background Tasks Agent

Tables

MF-Chairs

Lifter / Walkers Component Selection

Sensors

Hardware / Actuators

Fig. 2. Overview of the proposed approach

Mediator Agent This multi-purpose agent is responsible for accepting commands from the user interface’s Link Layer, initiating task execution procedures and querying for speciﬁc module information. Finally, after the task is executed, it informs the interface for changes. It also receives safety notiﬁcations from the Background Tasks Agent and initiates actions accordingly, giving them priority. System State Agent It is responsible for providing management and naming services for the rest of the system and monitoring the state of the system’s application modules. It informs, upon request, the rest of the system about the available modules and their state. The possible states of each application module depend on its functionality. For each robotic platform its state is a tuple < d, t > showing if its docked with an application module and which task it is currently undertaking. For an application module its state shows if it is docked with a platform. Task Allocation Agent If it is assigned a task by the Mediator Agent, then decomposes it, so that the necessary components are chosen and initiates a negotiation procedure similar to Contract Net Protocol, in order to ﬁnd the best module to perform the task. After it receives bids from the Component Agents it chooses a winner (the highest bidder) and assigns it the task. When it receives information from the Component Agent that the task was performed

20

G. Lidoris and M. Buss

it informs the Mediator. In the case of passive modules like tables and chairs we don’t have a task assignment but the winner is the module that is more appropriate for a given task. For example if we want a given platform to go and bring a table, then the negotiation ﬁnds the most suitable one (e.g. the one that is closer) and the Mediator is informed, so that it can assign the task to the platform. Background Tasks Agent Monitors system components and issues commands to the Mediator Agent, in order to deal with safety issues. For example, we want to avoid the risk of power failure of our mobile robotic platforms during task execution. Therefore the Background Tasks Agent constantly checks the power level of all mobile platforms. If a platform’s power level is found less than an acceptable minimum, then a Charge command is issued to the Mediator and the platform in question charges its batteries. It also can be used for setting out alerts for pre-programmed system maintenance operations, module failures etc. Generally the commands issued by this agent to the Mediator Agent have higher-priority as normal commands. Component Agents Each of these agents is created when a new hardware module is inserted to the system and it represents this module to the MAS. They are involved in the component selection negotiation procedure. If they are assigned a task they initiate a pre-programmed set of behaviours in order to accomplish it. The task description must contain information about its type and also other speciﬁc information regarding possible parts of the task. These agents have control over the hardware they embody and pass appropriate commands to it, by sequencing high level movement descriptions into low level commands. When the level of autonomy is set to manual, they just pass low level commands that are received from the UI to the hardware. Passive modules such as chairs, tables, etc, do not perform tasks themselves but need to be docked with a platform. They take part to the negotiations initiated by the Task Allocation Agent when we need to ﬁnd the most suitable module, in order to minimize resource usage. For example let’s suppose that the user wants a given platform to bring a table. Table Component Agents compute their utilities based on their distance from the platform. This is useful especially for home use scenarios, where only one platform exists but many tables and chairs, which are really inexpensive. Platform Component Agents rely on a behaviour based system. Behaviours are seen as condition / action pairs where conditions are logical expressions over the inputs and actions are themselves behaviours. In both cases, a behaviour is a directed acyclic graph of arbitrary depth. The leaves of the graphs are the behaviour type’s respective outputs. This notion of behaviour is also consistent with that laid out in [17] and [18]. Behaviours are nested at diﬀerent levels: selection among lower-level behaviours can be considered a higher-level behaviour, with the overall agent behaviour considered a single ”do-the-task” behaviour. These resulting highlevel tasks are also interconnected with each other and all together consist

A Multi-agent System Architecture for Modular Robotic Mobility Aids

21

our architecture’s task model. Figure 3 illustrates the interconnection between these tasks and how they are decomposed to simpler behaviours. robot+ module to target

robot to target

docked with module

no

yes

path planning no user confirmation

...

no

...

platform localization

dock with module

target reached yes stop

path execution

no

obstacle detection

yes

obstacle avoidance

no target reached

path tracking

yes stop

Fig. 3. Task model

When a robotic platform is assigned a task, it becomes unavailable for further tasks. This means that it does not take part in negotiations. Instead it monitors task execution and if no progress is made, it contacts the Task Allocation Agent to re-assign the task. Monitoring of task execution can simply be made by recalculating the utility of the task, in a given time space. If the utility remains unchanged or decreases below a given threshold, task execution must be terminated and the task has to be re-allocated. In the following section we will give a deﬁnition of task utility.

4 Utility for Task Assignment Component selection based on performance measures and task requirements is of great importance for our system. Imagine use cases where multiple modules of the same functionality exist. For example a user may have more than one table modules, which are really inexpensive, in his house. The system must then be in position to choose the best module for a given task e.g. in a fetch and carry scenario, where the mobile platform has to dock with a table, reach a goal location and return to the user, the best module in order to minimize resource usage would be the one closer to the goal location. We are dealing here with an instance of the Optimal Assignment Problem (OAP) [19] that was

22

G. Lidoris and M. Buss

originally studied in game theory. A formulation of this problem for robotics can be found in [20]. To address this problem we use the concept of utility. The idea is that each individual can internally estimate the value of executing an action. This estimation results a utility function which is highly domain and task dependent. An instructive deﬁnition of utility follows [11]. It is assumed that each agent / hardware module is capable of estimating its ﬁtness for every task of which it is capable. This estimation takes into account two task- and robot-dependent factors: • expected quality of task execution, given the method and equipment to be used • expected resource cost, given the spatio-temporal requirements of the task ( e.g. the length of the path to a target etc) So given a module M and a task T, if M is capable of executing T, then we can deﬁne QM T (p) and CM T (d) as the quality and cost, respectively, expected to result from the execution of T by M. Quality of task execution can be seen as a function of application module speciﬁc parameters, denoted by p. An example can be the power level of the battery of a robotic mobile platform. If the power levels are low then the platform will have to charge soon, therefore becoming temporary unavailable to execute tasks. Cost can be seen as a function of the path length from the goal position. A non-negative utility measure can now be deﬁned. UMT =

QMT (p) − CMT (d) if M is capable of executing T and QMT (p) > CMT (d) 0 otherwise

Each agent’s objective is to maximize utility. Utility maximization is equivalent in our case with eﬀective resource usage and therefore better system performance. We can incorporate learning into utility estimation, therefore adapting the system’s performance to the user behaviour and needs.

5 Experimental Design To validate our system’s functionality and take some ﬁrst simulation results, we have created an experimental testbed. For the implementation of our agents we decided to use JADE (Java Agent Development Framework) as a middleware. JADE is a software development framework aimed at developing multi-agent systems and applications conforming to FIPA (Foundation for Intelligent Physical Agents) standards for intelligent agents. Using JADE as a middleware basically means that we take advantage of its communication structures and naming directory services, in order to develop our experimental set-ups eﬀectively and with respect to the standards of the agent research community. This way we are able to analyse interactions between agents, as well as decision-making procedures. For the simulations we have used the

A Multi-agent System Architecture for Modular Robotic Mobility Aids

23

Player/Stage simulator. The structure of our experimental implementation is illustrated in Figure 4. From the discussion above it has been made evident that a large number of possible system states and setups can exist, making it almost impossible to simulate all of them. For that reason we had to choose some scenarios to be simulated which would test the task allocation and fault prevention mechanisms of our architecture. The following three were chosen: • one mobile robotic platform, two diﬀerent types of application modules and multiple tasks to be performed autonomously • multiple mobile robotic platforms, two diﬀerent types of application modules and one single task to be performed autonomously, with robot malfunction and task reallocation • multiple mobile robotic platforms, two diﬀerent types of application modules and multiple tasks to be performed autonomously Results from the simulations are very encouraging, with the system assinging tasks to modules in an eﬃcient manner. Even when task execution was interrupted on purpose by us, tasks were re-allocated succesfully. Agent communication is optimal posing almost no communication overhead. More specif-

JADE Platform UI-Agent

System State Agent

Mediator Agent

Task Deocomposition Agent

Platforms PlayerClient PositionPlayerDevice LaserPlayerDevice

Background Tasks Agent

Chairs PlayerClient PositionPlayerDevice LaserPlayerDevice

.....

Stage Player

GUI

Fig. 4. Experimental implementation

24

G. Lidoris and M. Buss

ically, agent interactions are shown in Figure 5 for the task of driving an initially undocked chair (ch1) to a given destination (target). FIPA Agent Communication Language performatives are used. UI-C / Link Layer

Mediator Agent

System State Agent

Task Allocation Agent

Background Tasks Agent

Component Agents

inform(ch1,auto,goto(target)) query-if(ch1,”docked”) inform(ch1,null) inform(”performing docking”) request(ch1,dock) cfp(ch1,dock) propose(bid) accept(pl1) inform(”done”) inform(”done”,pl1) inform(”done docking”) request(pl1,goto(target)) inform(”done”) inform(”done”)

Fig. 5. Agent interactions while performing a complex task

At ﬁrst Mediator Agent checks with the System State Agent whether the requested chair is docked with a mobile platform or not. The answer is negative, so a module selection procedure is iniated by the Task Allocation Agent in order to ﬁnd the most suitable platform (pl1 in Figure). When it is found, it is ordered to dock with the chair and then reach the deﬁned destination. The User Interface is appropriately updated by the Mediator while task execution is progressing. We can see that only few messages are used to accomplish a complicated task.

6 Conclusions We have deﬁned a Multi-Agent System architecture for modular mobility aids in terms of software architecture, modelling and requirements. We have shown how components interact with each other in order to achieve eﬀective functionality of the whole system. A quantiﬁcation of each module’s ability to undertake a task was also presented and used in a negotiation based module selection mechanism. Throughout this eﬀort, our main concern was simplicity

A Multi-agent System Architecture for Modular Robotic Mobility Aids

25

and functionality, given the safety critical nature of the application domain. Finally, a complete experimental testbed has been presented which has been used to validate our concepts. Since this is a new and demanding application domain much work remains on the architecture. We need to develop complete performance measures and enhance the system with learning mechanisms that will allow it to adapt its performance to the user’s behaviour. We also have to look deeper into mechanisms for fault prevention and system recovery. Finally, tests on physically embodied robots will be made.

References 1. Mandel C, Huebner K, Vierhuﬀ T (2005) Towards an Autonomous Wheelchair: Cognitive Aspects in Service Robotics. In Proceedings of Towards Autonomous Robotic Systems (TAROS 2005): 165-172 2. Gomi T, Griﬃth A (1998) Developing intelligent wheelchairs for the handicapped. In: Assistive Technology and Artiﬁcial Intelligence: Applications in Robotics, User Interfaces and Natural Language Processing 3. Hoyer H (1999) The OMNI-Wheelchair - State of the Art. In: Proceedings of Conference on Technology and Persons with Disabilities, Los Angeles 4. Simpson R et al (1998) Navchair: An assistive wheelchair navigation system with automatic adaptation. In: Assistive Technology and Artiﬁcial Intelligence: Applications in Robotics, User Interfaces and Natural Language Processing 5. Parker L (1998) Alliance: An architecture for fault tolerant multirobot cooperation. In: IEEE Transactions on Robotics and Automation, 14(2): 220-240 6. Simmons R, Smith T, Dias M B, Goldberg D, Hershberger D, Stentz A, Zlot R (2002) A layered architecture for coordination of mobile robots. In: Multi-Robot Systems: From Swarms to Intelligent Automata 7. Botelho S, Alami R (1999) M+: A schema for multi-robot cooperation through negotiated task allocation and achievement. In: Proceedings of the IEEE International Conference on Robotics and Automation: 1234-1239 8. Zlot R, Stentz A, Dias M B, Thayer S (2002) Multi-robot exploration controlled by a market economy. In: Proceedings of the IEEE International Conference on Robotics and Automation: 3016-3023 9. Gerkey B, Mataric M J (2002) Sold!: Auction methods for multi-robot coordination. In: IEEE Transactions on Robotics and Automation, 16(5):758-768 10. Dias M B, Zlot R M, Kalra N, Stentz A (2005) Market-Based Multirobot Coordination: A Survey and Analysis. Tech report CMU-RI-TR-05-14, Robotics Institute, Carnegie Mellon University 11. Gerkey B, Mataric M J (2004) A formal analysis and taxonomy of task allocation in multi-robot systems. In: Intl. Journal of Robotics Research, 23(9):939954 12. Smith R G (1980) The Contract Net Protocol: high level communication and control in a distributed problem solver. In: IEEE Transactions on Computers, C-29(12) 13. Gat E (1997)On three-layer architectures. In: Artiﬁcial Intelligence and Mobile Robots. MIT AAAI Press

26

G. Lidoris and M. Buss

14. Bonasso P, Kortenkamp D (1995) Characterizing an architecture for intelligent, reactive agents. In: Working Notes, AAAI Spring Symposium on Lessons Learned from Implemented Software Architectures for Physical Agents: 29-34 15. Gat E. (1991) Integrating planning and reacting in a heterogeneous asynchronous architecture for mobile robots. In: SIGART Bulletin 2 16. Connell J (1992) A hybrid architecture applied to robot navigation. In: Proceedings of IEEE ICRA ’92, Nice France 17. Stone P (1998) Layered learning in multi-agent systems. Phd Thesis, Carnegie Mellon University 18. Mataric M (1994) Interaction and intelligent behavior. Phd Thesis, MIT AI Lab 19. Gale D (1960) The theory of linear economic models. McGraw-Hill Book Company, New York 20. Gerkey B (2003) On multi-robot task allocation. Phd Thesis, University of South California, Los Angeles

How to Construct Dense Objects with Self-Reconﬁgurable Robots Kasper Stoy University of Southern Denmark, Campusvej 55, DK-5230 Odense M, Denmark [email protected] Summary. Self-reconﬁgurable robots can change their shape by rearranging the modules from which they are built. This self-reconﬁguration process has proven diﬃcult to control, because it involves control of a distributed system of autonomous, but mechanically coupled modules connected in time-varying ways. In this work we address this control problem and speciﬁcally look at how to construct dense conﬁgurations. The basic idea behind our solution is that holes inside the desired conﬁguration and modules outside attract each other. The eﬀect is that holes propagate to the surface of the desired conﬁguration and modules outside move in to ﬁll them. This basic solution is augmented with a scaﬀold and gradient-based algorithm which ensures that the robot does not get stuck in local minima or get disconnected during self-reconﬁguration. The approach is evaluated in simulation and we ﬁnd that the self-reconﬁguration process always converges and the time to complete a conﬁguration scales approximately linearly with the number of modules. This is, to the best of our knowledge, the ﬁrst time this result has been obtained for dense conﬁguration.

1 Introduction Reconﬁgurable robots are robots built from modules. These robots can be reconﬁgured by changing the way these modules are connected. If a robot is able to change the way the modules are connected autonomously, the robot is a self-reconﬁgurable robot. Self-reconﬁgurable robots are versatile because they can adapt their shape to ﬁt the task. Self-reconﬁgurable robots are also robust, because if a module fails, it can be ejected from the system and be replaced by a spare module. Potential applications for such robots include search and rescue missions, planetary exploration, building and maintaining structures in space, and entertainment. In order to realize these applications, research has to focus both on the development of the hardware of self-reconﬁgurable robots and on their control. The latter is the focus of this paper.

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 27–37, 2006. © Springer-Verlag Berlin Heidelberg 2006

28

K. Stoy

Fig. 1. This ﬁgure shows a simulated selfreconﬁgurable robot transforming from an initial random conﬁguration to a conﬁguration resembling a rhinoceros.

Fig. 2. This shows how a CAD model is represented using a set of overlapping bricks. This representation is used to drive the self-reconﬁguration algorithm.

One of the fundamental control problems of self-reconﬁgurable robots is control of the self-reconﬁguration process: the process by which the robot changes from one shape to another. An example of a self-reconﬁguration process is shown in Figure 1, where a simulated self-reconﬁgurable robot reconﬁgures from a random initial conﬁguration to a conﬁguration resembling a rhinoceros. The method proposed here consists of two steps. In the ﬁrst step a 3D CAD model, representing the desired conﬁguration, is transformed into a set of overlapping bricks as shown in Figure 2. The set of overlapping bricks is just a space eﬃcient way to represent the volume described by the CAD model. The second step is the actual self-reconﬁguration process. The desired conﬁguration is grown from a seed using the volume described by a set of overlapping bricks to guide the growth process. The growth process consists of two parallel-running processes: the scaﬀold and the ﬁll process. The scaﬀold process is responsible for growing a porous scaﬀold which guarantees convergence and connectivity of the system. The ﬁll process is responsible for ﬁlling in the pores of the scaﬀold and in this way create a dense conﬁguration. The solution to the self-reconﬁguration problem presented here is, to the best of our knowledge, the ﬁrst self-reconﬁguration algorithm which can be used to build dense conﬁgurations while only mildly restricting the class of conﬁgurations that can be constructed. At a more general level the algorithm we present provides a scalable, systematic, and convergent solution to the self-reconﬁguration problem.

How to Construct Dense Objects

29

The paper is structured as follows. In Section 2, the related work is described. In Section 3, we describe the algorithm we use to transform the 3D CAD model into a set of overlapping bricks. Section 4 describes how the representation of the desired conﬁguration interacts with the other components of the system to realize the self-reconﬁguration algorithm. Finally, in Section 6, the system is evaluated using a simulated self-reconﬁgurable robot where scalability and convergence properties are documented.

2 Related Work The self-reconﬁguration problem is: given a start conﬁguration, possibly a random one, how to move the modules in order to arrive at the desired ﬁnal conﬁguration. It is computational intractable to ﬁnd the optimal solution (see [4] for a discussion). Therefore, self-reconﬁguration planning and control are open to heuristic-based methods. Chirikjian, Pamecha, and Chiang propose iterative improvement as a way around the computational explosion [4, 9, 3]. The idea is to ﬁnd a suboptimal sequence of moves, which leads from the initial conﬁguration to the ﬁnal conﬁguration. This suboptimal sequence of moves can then be optimized using local searches. Rus et al. simplify the planning problem by using a speciﬁc intermediate chain conﬁguration that is easy to conﬁgure into and out of [11]. Kotay, Unsal, and Prevas propose a hierarchical planner [7, 15, 10]. At the highest level, some of the motion constraints of the underlying hardware are abstracted away to facilitate eﬃcient planning. Based on these high-level plans, the lower level then produces the detailed sequence of actions. Another approach, suggested by Kotay [7], is to use meta-modules consisting of a small number of modules. The idea is that by planning at the meta-module level there are no or few motion constraints. This makes the planning tractable. On the other hand, meta-modules consist of several modules, making the granularity of the robot larger. A related approach is to make sure that the robot maintains a uniform scaﬀolding structure, facilitating planning [14]. Butler implemented the distributed Pacman algorithm on the Crystalline robot [2]. The Pacman algorithm was improved and implemented on the Telecubes by Vassilvitskii [16]. These two robots have few motion constraints, making the planning problem easier. The approaches reviewed so far are deterministic and planning-based. The following approaches are not deterministic, but are much simpler, because a planning process is not needed. In early work, the idea was to use local rules as far as possible and then add randomness to deal with the problems that could not be solved using local rules. This is, for instance, true for the work on Fracta [8] and also later work on other robots [18]. The problem tended to be that even though the robot often ended up in the desired conﬁguration, it was not always so. We refer to this as the problem of convergence. This problem was also present in the work of Yim et al [17]. However, the chance of reaching

30

K. Stoy

the ﬁnal conﬁguration was higher in this work, because local communication was used to get some idea about the global shape at the local level. There are two solutions to the problem of convergence. One solution, proposed by Bojinov et al. [1], is not to focus on a speciﬁc conﬁguration. Instead, the idea is to build something with the right functionality. Using this approach it is acceptable if a few modules are stuck as long as the structure maintains its functionality. Alternatively, Jones et al. insist on a speciﬁc conﬁguration, but achieve convergence by enforcing a speciﬁc sequence of construction [6]. This paper is based on the work described in [12, 13], but extends this work by handling dense conﬁguration.

3 From CAD Model to Representation It is diﬃcult and time-consuming to hand-code local rules that result in a desired conﬁguration being assembled. Therefore, we need an automatic way of transforming a human-understandable description of a desired conﬁguration into a representation we can use for control. In our system, the desired conﬁguration is speciﬁed using a connected three-dimensional volume in the VRML 1997 or Wavefront .obj ﬁle format, which are industry standards produced by most CAD programs. We transform the model into a representation whose size scales with the complexity of the three-dimensional model. The representation consists of a set of overlapping bricks of diﬀerent sizes which approximates the input shape. The choice to use bricks is fairly arbitrary and other basic geometric shapes, such as spheres or cones, could be used as well. The representation is automatically generated from a CAD model. In the ﬁrst step, the algorithm builds an approximation of the CAD model using unit-bricks. A unit-brick is placed at a given coordinate, if its coordinate is contained in the model. This is true if a line from this coordinate to inﬁnity intersects the CAD model an uneven number of times. Afterwards, the algorithm optimizes the representation by repeatedly combining side-sharing bricks and by eliminating bricks which are contained in other bricks. Figure 3 shows a simple example of two overlapping bricks and their representation.

4 From Representation to Self-Reconﬁguration Algorithm The challenge is to develop a distributed control algorithm which can reconﬁgure the robot from an initial random conﬁguration to the desired conﬁguration as described by the representation. The basic idea behind our solution is that holes inside the desired conﬁguration and modules outside attract each other. The eﬀect is that holes propagate to the surface of the desired conﬁguration and modules outside move in to ﬁll them. The end result is that the volume

How to Construct Dense Objects

B

31

A

A: (0,0,1) → (3,1,2) B: (0,0,0) → (2,2,2)

Fig. 3. This ﬁgure shows two overlapping bricks and their representation (left). In this example eight unit-cubes are placed in the box labeled B and one in the box labeled A. These unit-cubes merge until only two cubes are left. One cube containing B with corner coordinates (0,0,0) and (2,2,2) and one containing A and overlapping B with corner coordinates (0,0,1) and (3,1,2). This ﬁgure also shows a basic unit of scaﬀold consisting of four modules (middle) and a lattice consisting of 8 scaﬀold elements (right). A lattice built from these scaﬀold elements is porous allowing modules to pass through it.

of the desired conﬁguration is ﬁlled implying that the desired conﬁguration has been reached. 4.1 Gradients Holes and modules use gradients to attract each other. Gradients provide modules with information about relative distances and directions in the conﬁguration. A gradient is created by a source module which sends out an integer, representing the concentration, to all neighbours. A non-source module calculates the concentration at its position by taking the minimum received value and adding one. This concentration is then propagated to all neighbours and so on. For example, if the source module sends out 100, neighbours of this module will have a concentration of 101, their neighbours 102 and so on. A module can locate the source by descending the gradient of concentration. However, if a module has to rely on the basic integer-based gradient to locate the source, it would have to move around randomly for a while to detect the direction of the gradient. In order to eliminate these unnecessary moves, we introduce a vector gradient which makes direction information available locally. The basic gradient implementation is extended with a vector indicating the local direction of the gradient. This vector is propagated by copying the vector from the neighbour with the lowest concentration and adding a unit vector in the direction of this neighbour and renormalizing the result. We use local message passing to implement the gradient. A message takes one time step to travel between neighbours and therefore it can take many time steps for gradients to be propagated in the system.

32

K. Stoy

4.2 Building a Porous Scaﬀold Unfortunately, the basic gradient-based algorithm alone does not work, because the algorithm also has to ensure that the robot does not get stuck in local minima or get disconnected during self-reconﬁguration. In order to address these problems we introduce a porous scaﬀold. In general, the set of positions contained in the scaﬀold Ps can be a subset of the positions in the set of positions in the desired conﬁguration Pdc . However, in our implementation we use a ﬁxed scaﬀold as shown in Figure 3. The algorithm that constructs the scaﬀold works as follows. All modules are initially connected in a random conﬁguration and have a copy of the representation of the desired conﬁguration. An arbitrary module is given an arbitrary coordinate contained in the desired conﬁguration and the scaﬀold. This ensures that the seed module is not placed in a hollow part of the conﬁguration. The seed module deﬁnes the coordinate system of the robot. Throughout the self-reconﬁguration process all modules keep track of their position in this coordinate system using local communication. Note that internal localization is a simple problem in self-reconﬁgurable robots, because all modules are physically connected throughout the self-reconﬁguration process as opposed to multi-robot systems where localization is a major challenge. If a module at random moves into a position next to a scaﬀold module and ﬁnds that this position is also part of the scaﬀold and the desired conﬁguration, it changes to the scaﬀold state. In this way the scaﬀold is grown from the ﬁrst scaﬀold module. The porous scaﬀold simpliﬁes the reconﬁguration problem, because it allows modules to move through the structure and thereby remove local minima from the conﬁguration. The scaﬀold is also used to prevent modules from disconnecting from the robot. Modules in the system cannot move independently of each other, because they depend on each other for connection to the robot. The problem is then to keep the system connected while allowing wandering modules to move. In our solution scaﬀold modules emit a connection gradient and modules only move if they do not perturb the gradient. This is enough to guarantee that the system stays connected. Detailed rules and proofs were presented in [12]. The scaﬀold puts a mild constraint on the class of conﬁguration where convergence is guaranteed. In our implementation the constraint is that the desired conﬁguration can be covered with a set of overlapping 2x2x2 cubes. Any cube needs an overlap of at least four modules on the same face with neighbouring cubes. This constraint ensures that there is one connected scaffold in the desired conﬁguration. This constraint prevents the algorithm from working with desired conﬁgurations containing sub-conﬁgurations that are one module thick. This constraint may be reduced or might even be removed if the scaﬀold is generated speciﬁcally for each conﬁguration.

How to Construct Dense Objects

33

4.3 Self-Reconﬁguration Algorithm We can now implement the self-reconﬁguration algorithm on top of these support algorithms. The idea here is, as mentioned, to 1) propagate holes toward modules outside the desired conﬁguration and 2) move modules outside the desired conﬁguration towards holes in the desired conﬁguration. A module can be in one of three states: wander, ﬁll, and scaﬀold. Wander modules are outside the desired conﬁguration. Fill modules are inside the desired conﬁguration, but are not part of the scaﬀold. Finally, scaﬀold modules are inside the desired conﬁguration and belong to the scaﬀold. Wander modules descend a hole gradient created by scaﬀold or ﬁll modules to reach holes in the desired conﬁguration. Wander modules also emit a wander gradient. Wander modules change to ﬁll when their positions are contained in the desired conﬁguration and are positioned next to a scaﬀold module. If, in addition, their positions are included in the scaﬀold, they change to scaﬀold. Fill modules climb the wander gradient, if and only if it does not take them out of the desired conﬁguration. This means that ﬁlling modules move away from wandering modules and in eﬀect propagate holes towards the surface and the wandering modules. Fill modules change to scaﬀold if they move into a position contained in the scaﬀold. Both ﬁll and scaﬀold modules emit the hole gradient if a neighbour position is unﬁlled and part of the desired conﬁguration. The self-reconﬁguration algorithm runs in parallel with the scaﬀold building algorithm. The two algorithms interact implicitly with each other: 1) the scaﬀold is porous and therefore allows the self-reconﬁguration algorithm to run. 2) the self-reconﬁguration algorithm moves modules to scaﬀold positions making the scaﬀold algorithm eﬃcient.

5 Simulated System In our simulation, we use modules, which are more powerful than any existing hardware platforms, but do fall within the deﬁnition of a Proteo module put forward by Yim et al. [17]. The modules are assumed to be cubical and when connected they form a lattice structure. They have six hermaphrodite connectors and can connect to six other modules in the directions: east, west, north, south, up, and down. Modules directly connected to a module are referred to as neighbours. A module can sense whether there are modules in neighbouring lattice cells. In this implementation, we do not control the actuator of the connection mechanism, but assume that neighbour modules are connected and disconnected appropriately. A module can only communicate with its neighbours. It is able to rotate around neighbours and to slide along the surface of a layer of modules. Finally, we assume that direction vectors can be uniquely transformed from

34

K. Stoy

the coordinate system of a module to the coordinate systems of its neighbours. This is necessary to propagate the vector gradient and use the brick representation. The simulator is programmed in Java3D. The simulation uses time steps. In each time step all the modules are picked in a random sequence and are allowed: 1) to process the messages they have received since last time step, 2) to send messages to neighbours (but not wait for reply), and 3) to move if possible. Note that this implies that it can take many time steps for gradients to be propagated in the system. In order for a module to move it has to satisfy the motion constraints of the system. This means that in order to move to a new position the new position has to be free. If using the rotate primitive, the position that the module has to rotate through to get to the new position has to be free as well. This is all assumed to be sensed locally. In our implementation the problem of two modules moving into the same cell at the same time is not addressed, because in the simulation only one module is allowed to move at a time. However, in a real parallel-running system this problem might be solved by rolling the module back to its original position in case of collision or through local right-of-way rules.

6 Experiments The task we pick for our experiment is to reconﬁgure the self-reconﬁgurable robot from a random initial conﬁguration to another random ﬁnal conﬁguration satisfying the constraint on desired conﬁgurations. The representations of the desired conﬁgurations are automatically generated. The representation is then downloaded into the modules of the simulation and the assembly process is started. We repeated the experiments ten times with changing numbers of modules. For each experiment we recorded the time steps needed, the total number of moves, and calculated the theoretical lower bound on the needed number of moves. The lower bound is the minimum number of moves needed to reconﬁgure from the initial conﬁguration, to the ﬁnal conﬁguration assuming that interference with other modules can be ignored [5]. The lower bound was only calculated for conﬁgurations with 800 modules or below, because with a higher number of modules it became computationally infeasible. The time steps needed to reconﬁgure grow sub-linearly with the number of modules (see Figure 4) which is comparable to our earlier results with porous conﬁgurations [12] and the results reported by Yim et al. [17]. Figure 4 shows that the number of moves grows approximately linearly with the number of modules. These results do of course depend on the shape of the desired conﬁgurations which, even though they are randomly generated, tend to be spherical for larger numbers of modules. In all experiments the system converges to the desired conﬁguration.

How to Construct Dense Objects 800

70000

700

60000

600

Moves

Time steps

Lower bound Observered

50000

500 400 300

40000 30000 20000

200

10000

100 0

35

0

500

1000

1500

2000

Modules

2500

3000

3500

0

0

500

1000

1500

2000

2500

3000

Modules

Fig. 4. This ﬁgure shows how the number of time steps (left) and moves (right) needed to reconﬁgure depends on the conﬁguration and the number of modules. The theoretical lower bound on the number of moves is also shown.

Fig. 5. This ﬁgure shows the total number of messages used for communication as a function of the number of modules. It is shown how the messages are divided between diﬀerent tasks (the order of items in the legend is the same as in the column stacks).

The number of moves and the time to complete a reconﬁguration are important characteristics of an algorithm. However, another aspect is the amount of communication needed. This information is summarized in Figure 5. The system relies heavily on communication and in case of the 3200-modules experiments on average 8.9 ± 2.3 messages are sent per module per time step. These messages originate from propagating three changing gradients, propagating position information, and message exchanges need to run the distributed connection algorithm. In our current implementation, little is done to minimize the number of messages, so there is room to improve the performance.

36

K. Stoy

7 Conclusion and Future Work We have presented an approach to self-reconﬁguration working in two steps. First, a representation of a desired conﬁguration is generated automatically based on a 3D CAD model. Second, a self-reconﬁgurable robot uses this representation to transform itself from a random initial conﬁguration to the desired conﬁguration. The transformation is controlled by an algorithm which propagates holes in the desired conﬁguration towards modules outside. This algorithm is extended with a scaﬀold building algorithm that removes local minima from the desired conﬁguration and thereby guarantees convergence of the algorithm. Furthermore, a connection algorithm ensures that the robot stays connected during the transformation process by guaranteeing that modules at all times are connected to the scaﬀold. In experiments with a simulated self-reconﬁgurable robot we have demonstrated that the system represents an eﬃcient, systematic, and convergent solution to the self-reconﬁguration problem, while only restricting the class of possible conﬁgurations mildly. The algorithm represents, to the best of our knowledge, the ﬁrst algorithm able to maintain these characteristics while handling dense conﬁgurations.

Acknowledgments The author would like to thank George Konidaris, University of Massachusetts at Amherst, for the fruitful discussions resulting in this paper.

References 1. H. Bojinov, A. Casal, and T. Hogg. Emergent structures in modular selfreconﬁgurable robots. In Proc., IEEE Int. Conf. on Robotics & Automation (ICRA’00), volume 2, pages 1734 –1741, San Francisco, California, USA, 2000. 2. Z. Butler, S. Byrnes, and D. Rus. Distributed motion planning for modular robots with unit-compressible modules. In Proc., IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS’01), volume 2, pages 790–796, Maui, Hawaii, USA, 2001. 3. C.-J. Chiang and G.S. Chirikjian. Modular robot motion planning using similarity metrics. Autonomous Robots, 10(1):91–106, 2001. 4. G. Chirikjian, A. Pamecha, and I. Ebert-Uphoﬀ. Evaluating eﬃciency of selfreconﬁguration in a class of modular robots. Robotics Systems, 13:317–338, 1996. 5. G.S. Chirikjian and J.W. Burdick. The kinematics of hyper-redundant robot locomotion. IEEE transactions on robotics and automation, 11(6):781–793, 1995. 6. C. Jones and Maja J. Matari´c. From local to global behavior in intelligent self-assembly. In Proc. of the IEEE Int. Conf. on Robotics and Automation (ICRA’03), pages 721–726, Taipei, Taiwan, 2003.

How to Construct Dense Objects

37

7. K. Kotay and D. Rus. Algorithms for self-reconﬁguring molecule motion planning. In Proc., IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS’00, volume 3, pages 2184–2193, Maui, Hawaii, USA, 2000. 8. S. Murata, H. Kurokawa, and S. Kokaji. Self-assembling machine. In Proc., IEEE Int. Conf. on Robotics & Automation (ICRA’94), pages 441–448, San Diego, California, USA, 1994. 9. A. Pamecha, I. Ebert-Uphoﬀ, and G.S. Chirikjian. Useful metrics for modular robot motion planning. IEEE Transactions on Robotics and Automation, 13(4):531–545, 1997. 10. K.C. Prevas, C. Unsal, M.O. Efe, and P.K. Khosla. A hierarchical motion planning strategy for a uniform self-reconﬁgurable modular robotic system. In Proc., IEEE Int. Conf. on Robotics and Automation (ICRA’02), volume 1, pages 787–792, Washington, DC, USA, 2002. 11. D. Rus and M. Vona. Self-reconﬁguration planning with compressible unit modules. In Proc., IEEE Int. Conf. on Robotics and Automation (ICRA’99), volume 4, pages 2513–2530, Detroit, Michigan, USA, 1999. 12. K. Støy. Controlling self-reconﬁguration using cellular automata and gradients. In Proc., 8th int. conf. on intelligent autonomous systems (IAS-8), pages 693– 702, Amsterdam, The Netherlands, 2004. 13. K. Støy and R. Nagpal. Self-reconﬁguration using directed growth. In Proc., int. conf. on distributed autonomous robot systems (DARS-04), pages 1–10, Toulouse, France, 2004. ¨ and P.K. Khosla. A multi-layered planner for self-reconﬁguration of a 14. C. Unsal uniform group of i-cube modules. In Proc., IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS’01), volume 1, pages 598–605, Maui, Hawaii, USA, 2001. ¨ 15. C. Unsal, H. Kiliccote, and P.K. Khosla. A modular self-reconﬁgurable bipartite robotic system: Implementation and motion planning. Autonomous Robots, 10(1):23–40, 2001. 16. S. Vassilvitskii, M. Yim, and J. Suh. A complete, local and parallel reconﬁguration algorithm for cube style modular robots. In Proc., IEEE Int. Conf. on Robotics and Automation (ICRA’02), volume 1, pages 117–122, Washington, DC, USA, 2002. 17. M. Yim, Y. Zhang, J. Lamping, and E. Mao. Distributed control for 3d metamorphosis. Autonomous Robots, 10(1):41–56, 2001. 18. E. Yoshida, S. Murata, H. Kurokawa, K. Tomita, and S. Kokaji. A distributed reconﬁguration method for 3-d homogeneous structure. In Proc., IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS’98), volume 2, pages 852–859, Victoria, B.C., Canada, 1998.

In Situ Autonomous Biomechanical Characterization Mehdi Boukallel, Maxime Girot, and St´ephane R´egnier Laboratoire de Robotique de Paris Universit´e Pierre et Marie Curie - Paris 6 CNRS - FRE 2507 18 route du Panorama - 92265 Fontenay Aux Roses - France boukallel, girot, [email protected] Summary. This paper presents a fully automated microrobotic system based on force/vision referenced control designed for cell mechanical characterization. The design of the prototype combines Scanning Probe Microscopy (SPM) techniques with advanced robotics approaches. As a result, accurate and non-destructive mechanical characterization based on soft contact mechanisms are achieved. The in vitro working conditions are supported by the experimental setup so that mechanical characterizations can be performed in biological environmental requirements as well as in cyclical operating mode during several hours. The design of the diﬀerent modules which compose the experimental setup are detailed. Mechanical cell characterization experiments under in vitro conditions on human adherent cervix Epithelial Hela cells are presented to demonstrate the viability and eﬀectiveness of the proposed setup.

Keywords: In vitro mechanical cell characterization; Scanning Probe Microscopy (SPM) techniques; human adherent cervix Epithelial Hela cells mechanical characterization.

1 Introduction Nowadays robotics and microrobotics techniques play an important role for exploring the mechanical cell behaviour. The ability to study accurately the mechanical behaviour of individual cells is a key component in understanding the elementary biological functions that various biological cells exhibit. Furthermore, the individual mechanical cell characterization is a major step towards the knowledge and understanding the behavior of the complex mechanical tissues. Up to date, several experimental setups have been developed to identify the control mechanisms of the cell mechanical response [1]-[5]. Among these robotics and microrobotics systems, the most promising ones

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 39–50, 2006. © Springer-Verlag Berlin Heidelberg 2006

40

M. Boukallel, M. Girot, and S. Regnier

involves Scanning Probe Microscopy (SPM) techniques for nanoscale. These techniques have the potential to give an accurate quantitative information about local forces and contact mechanisms. The Atomic Force Microscope (AFM) has become a commonly used tool to measure mechanical properties of the biological samples [6]-[10]. In these studies, a ﬂexible cantilever with low spring constant and an atomic sharp tip is brought in the vicinity of the biological sample. Deﬂection of the cantilever as a result of the mechanical interaction between the microindenter and the sample is monitored by a split photodiode and the use of a laser beam reﬂected on the back of the cantilever. The signiﬁcant research involving AFM make possible the measuring of relevant cells mechanical properties (Young’s modulus, bulk modulus, surface roughness, ...). However, most of these studies have not been performed in biological clean room environment. Since elementary biological functions and mechanical properties of biological cells are widely aﬀected by the experimental conditions, identiﬁed properties may not be relevant. Furthermore, as the reaction of the biological samples to stress vary greatly in a given lapse time, it is important to monitor the characterization process continuously in an in vitro environment. Moreover, some problems are commonly associated with the use of standard commercial cantilevers with a sharp tip for the soft sample mechanical characterization. The nanometers size dimensions of tips can cause important local strains which are higher than elastic domain. Furthermore, depending on the magnitude of the applied force on the soft samples, cantilevers tips as well as the samples can be easily damaged so that the local strain applied in the indented area becomes changed. To overcome these problems, Mahaﬀy et al. [11] suggest to attach a micro-sphere at the cantilever tip. However, besides the necessity of dexterous manipulations for accurate placement of the microsphere, this method needs to use well-know materials (geometry, mechanical properties, ...). Another diﬃculty is associated with the mechanical characterization of biological samples using sharp cantilever. Usually, the spectroscopy curves collected with the AFM are used in conjunction with an appropriate analytical model to estimate Young’s modulus, friction, wear and other material properties. According to the literature, the Hertzian model which describes the relationship between force and identation is the commonly approach used for ﬁtting the experimental data. Also, two major assumptions are made : linear elasticity and inﬁnite sample thickness. Some authors have showed that in the case of soft contact mechanisms, models derived from linear elasticity can lead to signiﬁcant errors [12]. Moreover, due to the imperfections of the tip radius of curvature an ill-deﬁned contact region results between the probe and the sample. Consequently, uncertainties are introduced for choosing the appropriate ﬁtting analytical model. It has been also shown, depending on the applied force and on the sample thickness, that large errors may result when using inﬁnite thickness models [13][14]. The authors compute force-displacement

In Situ Autonomous Biomechanical Characterization

41

curves for ﬁnite sample thickness to show that, for soft and thin sample the error in the estimated elasticity modulus can be an order of magnitude. In our opinion, the cell mechanical characterization in in vitro environment using tipless cantilever seems to be a promising solution. As study involving such cantilever are less prone to problems associated to sharp tip cantilever, enhanced non-destructive cell mechanical characterization should be achieved. For this purpose, a Force Bio-Microscope (FBM) system has been developed which combines SPM techniques and advanced robotics approaches. A tipless chemically inert cantilever is used for this study. Contact mechanisms resulting from the biomechanical process are modeled with the DMT (Derjaguin, Muller and Toporov) contact theory which taking into account both adhesion forces and ﬁnite sample thickness. In order to demonstrate the accuracy of the DMT model in the case of soft contact mechanisms, the estimated force-deformation curve is compared to the one predicated by the Hertz theory.

2 Experimental Setup Overview The automated Force Bio-Microscope (FBM) device is an hybrid AFM microscope associating both scanning microscopy approach and biological environment constraint. The FBM consists mainly of three units: the mechanical sensing unit which performs detection, positioning and sensing features, the imaging/grabbing unit for imaging and cell tracking features and the clean room in vitro unit which allows experiments to be conducted in biological environment. As the FBM experimental setup provides suitable conditions for study in controlled environment, the mechanical measurements process can be done on the biological sample on an extended lapse of time. A master computer is used to drive the FBM in a automatic operating mode based on force/vision referenced control. A user-deﬁnable graphical interface has been developed in order to make conﬁguration of the automatic cell mechanical characterization process easier. To avoid undesired mechanical vibrations during the cell characterization tasks, the FBM is installed on an anti-vibration table. The overall conﬁguration of the FBM and the diﬀerent working components are shown in ﬁgure 1(A). 2.1 Mechanical Sensing Unit The mechanical sensing unit is based on the detection of the deﬂection of a cantilever by an optical technique. A four quadrant photodiode with internal ampliﬁers associated to a low power collimated laser diode (wavelength 650 nm) are used in order to perform both axial and lateral nanoNewtons forces measurements. The optical path of the Gaussian laser beam is optimized using a pair of mirrors and a glass aspheric condenser lens. The total

42

M. Boukallel, M. Girot, and S. Regnier

sensing area of the photodiode is 7 mm2 with a spectral response from 400 to 1100 nm. The sensitivity of the optical detection device is 5 mV /µm. A low spring constant (0.2 N/m) uncoated tipless silicon cantilever is used as probe for the cell mechanical characterization. The lever is 450 µm long, 90 µm large and 2 µm thick. The biological samples are accurately positioned below the cantilever by a 3 DOF’s (x,y and z) micropositioning encoded stage with a submicrometer resolution (0.1 µm). The kinematics features of the micropositioning stage allows to achieve accurate mechanical measurements in a workspace of 25 x 25 x 25 mm3 with a good repeatability. The conﬁguration of the mechanical sensing unit including the optical detection device is presented in ﬁgure 1(B). A magniﬁed picture of the cantilever with the focused laser beam on its reﬂective surface is shown in the same ﬁgure.

(A) Cage incubator

100 mm

(B) 3 DOF micropositioning stage

Heating module

Antivibration table Inverted microscope

Miror

Laser diode

15 mm 10 mm Photodiode

Glass aspheric condenser lens Silicon tipless cantilever

CCD camera

10 mm

Fais

Focused laser beam 100 µm

Cantilever

Fig. 1. (A) The FBM experimental setup overview. (B) The mechanical sensing unit.

For the preliminary study, we focused on force feedback control of cantilever ﬂexural deﬂection. Thus, only the vertical micropositioning stage is servoed. By knowing the micromotors vertical position as well as deﬂection of the cantilever using the optical detection device, a optimized Proportional and Derivative (PD) controller was designed to ensure optimal control performance. The (PD) terms are optimized using the Ziegler-Nichols method. Figure 2(A) shows experimental results on the force referenced control approach for diﬀerent desired forces. 2.2 Imaging/Grabbing Unit The imaging/grabbing unit consists of an inverted microscope (Olympus IMT2) with Nikon 10x and 20x objectives. A phase contrast device is mounted on

In Situ Autonomous Biomechanical Characterization

43

the microscope for precise contrast operation. The inverted microscope is ﬁtted out with a CCD camera (754x488 pixels resolution). Using a frame grabber and a specialized imaging library package (Matrox Imaging) associated to the CCD camera, automatic mechanical characterization based on image features tracking is achieved (cf. ﬁgure 2(B)). The pixel to real world calibration of the CCD camera is achieved by means of a calibrated glass micro-array as well as calibrated micro-spheres (cf. ﬁgure 3). (A)

(B)

Cell target

100 µm

Cantilever

Fig. 2. (A) The force feedback control approach. (B) The vision feedback control approach.

Fig. 3. Pixel to real world calibration of the CCD camera using the calibrated glass micro-array.

44

M. Boukallel, M. Girot, and S. Regnier

2.3 Clean Room in vitro Unit The incubating system is formed with a controlled heating module which maintains temperature on the cage incubator at 37 o C using a single thermocouple probe. The cage incubator ensures a temperature stability within the 0.1 o C. A mixed stream composed by a 5% CO2 and humidiﬁed air is fed into a small incubating chamber containing the biological samples, avoiding in this way condensation on the cage walls that could damage the mechanical parts of the microscope and the micropositioning stage. The whole system including the FBM is placed in a positive pressure clean room to protect the biological environment.

3 In vitro Mechanical Characterization Experiments Experiments presented in this paper are focused on the mechanical stress response of human cervix Epithelial Hela (EpH) cells under controlled environment. In order to be valid, some assumptions are made in this regard: (i) the biological membrane is assimilated to a perfectly elastic spherical material; (ii) the enclosing liquid by the biological membrane of the EpH cells is considered as incompressible; (iii) the surfaces are assumed frictionless so that only normal forces are transmitted; (iv) the osmotic inﬂuence on volume modiﬁcation is neglected. The EpH cells can be assimilated morphologically to an elliptical cells with a thin surrounding biomembrane. In the present study, the average dimensions of the biological sample is 10 µm long, 9 µm large and 6 µm height (cf. ﬁgure 4). The Epithelial Hela cells (EpH) are prepared on Petri dishes with speciﬁc culture medium formed by Dulbecco’s Modiﬁed Eagle’s Medium (DMEM) with high glucose and L-glutamine components and 10 % of foetal bovine serum.

Fig. 4. (A) Magniﬁed image of the cervix Epithelial Hela cells obtained with an 63x objective. (B) The cervix Epithelial Hela cells morphology observed by ﬂuorescence techniques.

In Situ Autonomous Biomechanical Characterization

45

3.1 Cell’s Mechanical Response Characterization Figure 5(A) shows the experimental curves of the photodiode output as function of the sample displacement (Δz) performed on both single EpH cell and hard surface. The single step of the sample displacement is 200 nm and the total displacement is 8 µm. Deformations δ of the EpH cell are monitored by calculating the diﬀerence between the sample displacement Δz and the cantilever deﬂection Δd. The non-linear elastic behaviour of the EpH can be seen in the ﬁgure 5(B) which presents the sample deformation δ as function of the load force applied by the cantilever. (A)

(B)

3.5

Hard surface Biological sample

3

0.25

2.5 Deformation δ (µm)

Vertical photodiode output (V)

0.3

0.2

0.15

2

1.5

0.1

1

0.05

0.5

0 0

5 10 Sample displacement µm) (

0 0

0.1 0.2 0.3 Load forceµN) (

0.4

Fig. 5. (A) Experimental data of the photodiode output as function of the sample displacement performed on both single EpH cell and hard surface. (B) Experimental curve of the sample deformation δ as function of the applied load by the cantilever.

3.2 Viscoelastic Behaviour Characterization The viscoelastic behaviour of the EpH cells are also investigated by the FBM device. Cyclical and automatic approach and retract experimentations were conducted on the same biological sample during 2 hours with 3 minutes intervals. For this given study, the motion amplitude and the single step of the vertical microstage are ﬁxed to 8 µm and 200 nm, respectively. In order to reduce the cantilever damping oscillations during the mechanical characterization process, velocity of the sample positioning stage is chosen small (0.5 µm/s). Figure 6(A) shows 3 approaches and retracts curves monitored at different time intervals (t=0 mn, 40 mn and 80 mn) of the cyclical experiments. A single referenced approach and retract curves performed on hard surface are given in ﬁgure 6(B). According to the collected data, the EpH sample exhibit the same viscoelastic behaviour during all the experimentation.

46

M. Boukallel, M. Girot, and S. Regnier (A)

0.45

Appoach

t=40 mn

0.4

(B)

0.45

t=0 mn

Retract

0.4

t=80 mn 0.35

Measured force (µN)

Measured force (µN)

0.35

0.3

0.25

0.2

0.15

0.3

0.25

0.2

0.15

0.1

0.1

0.05

0.05

0

0 0

2

4

6

8

0

Sample displacement (µm)

2

4

6

8

Sample displacement (µm)

Fig. 6. (A) Experimental spectroscopy curves (approach and retract) performed on a single EpH cell at diﬀerent time intervals (t=0 mn, 40 mn and 80 mn). (B) Single referenced approach and retract curves performed on hard surface. 0.7

ac

t

0.6

Strain ε

Re

tr

0.5

0.4

Ap

ch

oa

pr

0.3

Adhesion forces

0.2

0.1

0

0

1

2

3

4

Sample displacement (µm)

5

6

Fig. 7. Dimensionless strain per module length ε as function of the sample displacement Δz.

In Situ Autonomous Biomechanical Characterization

47

Compared to the approach and retract curves performed on hard surface (cf. ﬁgure 6(B)), the retract collected data performed on the Eph sample are diﬀerent from the approach one. This suggests adhesion forces between the cantilever probe and the biological sample which modiﬁes the total strain energy. Figure 7 shows the strain ε according to the sample displacement Δz. This ﬁgure emphasizes adhesion interaction between the cantilever and the biological sample.

3.3 In vitro Eﬃciency Approach for Cell Mechanical Characterization In order to study correlation between the mechanical cell properties and the environmental culture conditions, we have experimented automatic and cyclical spectroscopy tasks on a single EpH cell during several minutes without the use of the incubating system. As the precedent study, the sample displacement and the single step of the vertical micropositioning stage are ﬁxed to 8 µm and 200 nm respectively. Figure 8(A) shows evolution of the EpH cell mechanical behaviour of cyclical spectroscopy operation with and without the use of the incubating system. More speciﬁcally, curve (A) shows the approach and retract curves using the cage incubator. Curves (B), (C) and (D) show the mechanical behaviour of the studied EpH cell for diﬀerent elapsed times t0 once the cage incubator is turned oﬀ. These mechanical characterization experiments obviously reveal that mechanical properties of the studied sample are aﬀected by the temperature conditions. This diﬀerence suggests that the intra or extra-cellular matrix react to the variation of temperature. 3.4 Analytical Model for Sample Deformation Estimation The deformation δ resulting from the mechanical cell characterization process is estimated using an appropriate analytical model. The DMT (Derjaguin, Muller and Toporov) theory is chosen for this purpose. Noting R the radius of the biological sample (R=10 µm), w the adhesion work, P the load force applied by the cantilever and a0 the radius of the contact area when P = 0, both contact area radius a and deformation δ can be expressed according to DMT [15] by a3 = a30 ( δ=

a2 R

P + 1) 2πRw

(1) (2)

48

M. Boukallel, M. Girot, and S. Regnier

The DMT model suggests that adhesion work w can be expressed according to the pull-oﬀ force Pof f needed to overcome adhesion forces as [15] Pof f = 2πRw

(3)

As the pull-oﬀ force Pof f and the contact area a0 are accurately measured using the FBM (Pof f 20 nN , a0 2 µm), the values of w and a0 are introduced in equation 1 for the determination of a. Equation 2 gives the estimation of δ. Figure 8(B) shows the estimation of the biological sample deformation δ as a function of the simulated load force P using the Hertz and the DMT theories. These analytical results are compared to the experimental measurements performed with the FBM and presented in section 3.1. As adhesion forces are not considered, large errors are observed between the experimental data and the predicated force-deformation curves (in order of 0.2 µm of magnitude) in the case of the Hertz model. (A)

3.5

0.4 t0 = 0 mn

(A)

t0 = 5 mn

0.3

t0 = 9 mn

0.25

t0 = 13 mn

(B)

2.5

(C)

0.2 (D)

0.15 0.1 0.05

2

1.5 1

0.5

0 0.05 0

3

Deformation δ (µm)

Measured force (µN)

0.35

(B) Experimental data Hertz DMT

1

2

3

4

5

6

Sample displacement (µm)

7

8

0

0

0.05

0.1

0.15

0.2

0.25

Applied load P (µN)

0.3

0.35

Fig. 8. Evolution of the measured force as a function of the sample displacement for diﬀerent elapsed times t0 = 0, 5, 9 and 13 mn. (B) Estimation of the biological sample deformation δ as a function of the simulated load force P using the Hertz and the DMT theories compared to the experimental data.

4 Conclusion This paper has presented the development of a micro-force sensing system for in vitro mechanical cell characterization. The experimental setup combines Scanning Probe Microscopy (SPM) techniques with advanced robotics approaches. As the developed system operates in a fully automatic mode based on visual and force tracking control, eﬀective mechanical characterization and reliable data acquisition are achieved. The Force Bio-Microscope

In Situ Autonomous Biomechanical Characterization

49

device (FBM) consists of three units with autonomous force sensing and measurements capabilities. Each unit is designed, calibrated or conﬁgured towards an eﬀective in vitro mechanical cell characterization. Mechanical characterization is conducted using the FBM on human adherent cervix Epithelial Hela cells. These experiments demonstrate the eﬃciency of the experimental setup developed to explore the mechanical properties in in vitro conditions of adherent biological samples. The contact mechanisms resulting form the cell mechanical characterization process are predicted using appropriate model (DMT) taking into account both adhesion forces and ﬁnite sample thickness.

5 Future work Current work involves modelling the soft contact mechanisms using a tipless cantilever taking into account friction, osmotic inﬂuence and anisotropic non-linear elastic properties. Future work will be focused on exploring the mechanical transduction of living cells in in vitro environment conditions. One of the most promising results of this study will be the description of both mechanosensitivity and mechanical transduction mechanisms at the microscale for biological cells or tissue conﬁguration.

Acknowledgment The authors are grateful to J. Angulo-Mora from the Genetic of the Radiosensitivity Laboratory (LGR) of the French Atomic Energy Commission (CEA), for cells preparation and for his enthusiastic help.

References 1. Sun Y, Nelson B J (2004) Jrnl. of Information Acquisition 1:23–32 2. Van Vliet K J, Bao G, Suresh S (2003) Acta Materialia journal 51:5881–5905 3. Kim D H, Yun S, Kim B (2004) in proc.of the Int. Conference on Robotics and Automation 5013–5018 4. Guck J, Ananthakrishnan R, Cunningham C C, Kas J (2002) J.Physics:Condens. Matter 14:4843–4856 5. Sharp J M, Clapp A R, Dickinson R B (2003) Colloids and Surface B:Biointerfaces 27:355–364 6. Radmacher M (1997) IEEE Engineering in Medicine and Biology 47–57 7. Sahin O, Yaralioglu G, Grow R, Zappe S F, Atalar A, Quate C, Solgaard O (2004) Sensors and Actuators A 114:183–190 8. Park S, Costa K, Ateshian G (2004) Journal of Biomechanics 37:1687–1697 9. Dimitriadis E K, Horkay F, Maresca J, Kachar B, Chadwick R S (2002) Biophysical Journal 82:2798–2810

50

M. Boukallel, M. Girot, and S. Regnier

10. Yao X, Walter J, Burke S, Stewart S, Jericho M H, Pink D, Hunter R, Beveridge T J (2002) Colloids and Surface B:Biointerfaces 23:213–230 11. Mahaﬀy R, Shih C K, Mackintosh F C, Kas J (2000) Phys. Rev. Lett. 85:880– 883 12. Costa K, Yin F C (1999) J. Biomeh. Engr. Trans. ASME 121:462–471 13. Domke J, Radmacher M (1998) Langmuir 14:3320–3325 14. Akhremitchev B, Walker G (1999) Langmuir 15:5630–5634 15. Maugis D (2000), Contact, Adhesion and rupture of elastic solids, SpringerVerlag

Incremental Learning of Task Sequences with Information-Theoretic Metrics∗ Michael Pardowitz, Raoul Z¨ ollner, and Rudiger Dillmann Institute of Computer Science and Engineering Universit¨ at Karlsruhe (TH) Karlsruhe, D-76138, Germany, P.O. Box 6980 Email: {pardow, zoellner, dillmann}@ira.uka.de Summary. Learning tasks from human demonstration is a core feature for household service robots. Task knowledge should at the same time encode the constraints of a task while leaving as much ﬂexibility for optimized reproduction at execution time as possible. This raises the question, which features of a task are the constraining or relevant ones both for execution of and reasoning over the task knowledge. In this paper, a system to record and interpret demonstrations of household tasks is presented. A measure for the assessment of information content of task features is introduced. This measure for the relevance of certain features relies both on general background knowledge as well as task-speciﬁc knowledge gathered from the user demonstrations. The results of the incremental growth of the task knowledge when more task demonstrations become available is demonstrated within the task of laying a table.

1 Introduction During the last years, human-centered robotics became a major trend in the robotics community. One crucial point in HCR is to design systems that adapt to unstructured and new situations the human user comes up with. This can be achieved by giving the robot the ability of learning to transform the knowledge the user has into knowledge utilizable by a mobile service robot. One of the most intuitive ways to acquire new task knowledge is to learn it from the human user via demonstration and interaction. This approach to task learning is widely known as Programming by Demonstration (PbD). PbD-systems should be capable of learning a task from a single demonstration suﬃciently for execution while being able to generalize it when further ∗

This work has been partially conducted within the german collaborative research center ”SFB Humanoide Roboter” granted by Deutsche Forschungsgemeinschaft, and within the EU Integrated Project COGNIRON (”The cognitive robot companion”), funded by the European Commision Division FP6-IST Future and Emerging Technologies under Contract FP6-002020.

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 51–63, 2006. © Springer-Verlag Berlin Heidelberg 2006

52

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

demonstrations of the same task become available. Incremental learning approaches that gradually reﬁne task knowledge pave the way towards suitable PbD-systems for humanoid robots. One aspect that can only be learned incompletely from a single user demonstration is the sequence the subparts of a certain task can be scheduled. So the task knowledge must explicitly encode those in order to ensure a reliable, faultless and save execution of the task. After seeing a single demonstration, PbD-systems can state multiple valid hypotheses on the sequential constraints a task must obey. When more demonstrations become available, the incorrect hypotheses conﬂicting with the new demonstrations can be pruned. Identifying the corresponding subtasks performed in diﬀerent demonstrations of the same task is not an easy job. Many features of the task are exposed to noise. Additionally, some features posess higher relevance to a certain task, while being negligible for others. A method for identifying the features that characterize a task and basing the recognition of corresponding subtasks on this selection is presented in this paper. The remainder of this paper is organized as follows: The next section gives an overview on related work concerning programming by demonstration and task learning from user demonstrations. Section 3 describes the system for the acquisition of task knowledge from a single user demonstration. Section 4 introduces task precedence graphs and proposes a method for learning the underlying precedence graph of a task with two or more sequential demonstrations given. The methods for identifying corresponding subtasks within diﬀerent task demonstrations described in section 5 are an important preliminary for this computation. Finally, the methods desscribed in earlier sections are evaluated in section 6.

2 Related Work Since [7] proposed a programming system for industrial robots using visual observation of a human user, several programming by demonstration (PbD) systems and approaches based on human demonstrations have been proposed during the past years. Many of them address special problems or a special subset of objects only. An overview and classiﬁcation of the approaches can be found in [5, 14]. The mentioned Systems can be discriminated by the learning paradigm applied: Imitation learning systems are used to re-identify and execute human motions [3, 1, 2], aiming at a high similarity of the robot’s movements to the demonstrated trajectories. Examples for the entities learned by imitation are skills like grasp or release actions ([15]) or moves of manipulated objects with constraints like ”move along a table surface” etc. The classical ”peg in hole” problem can be learned by imitation, for example using a force controller which minimizes a cost function or a task consisting of a sequence of skills like ”ﬁnd the hole”, ”set perpendicular position” and ”insert”. These systems

Incremental Learning of Task Sequences

53

require a large set of task demonstrations for generalizing the trajectory before the learning process can start. Background knowledge based or deductive PbD-systems, as presented in [13, 16], usually require much less or even only a single user demonstration to generate executable task descriptions. Often, the analysis of a demonstration takes place observing the changes in the scene, described by using relational expressions or contact relations [7, 10]. When mapping the learned task to the executing system, background knowledge based approaches are used in order to replicate the eﬀects in the environment [12]. Sequential analysis of tasks is presented in [4, 9]. The ﬁrst records multiple user demonstrations of a complex manipulation skill. Unnecessary actions that do not appear in all demonstrations are pruned and only the sequence of essential actions is retained. The latter stresses the role of interaction with the human user to facilitate learning of sequential arrangement of behaviors in a navigation task. In this paper a system is proposed that combines background knowledge based learning techniques with the subsymbolic, data-driven approaches into a single framework for learning of and reasoning on user task demonstrations . This gives the system to use the advantages of both learning paradigms. It overcomes the data-driven approaches’ need for a lot of demonstrations being available before the learning process can start, while being able to incrementally reﬁne the task knoweldge that is generated using background knowledge based methods. It is applied to analyse the sequential structure of a task. In advance to the method of [4], it can not only prune the unnecessary operations that do not contribute to the eﬀects of the demonstration, but reorder the remaining manipulation segments. This is done by applying symbolical reasoning methods on the diﬀerent demonstrations of a single task in an autonomous way, in contradition to the approch in [9].

3 A System to Generate Hierarchical Task Descriptions User demonstrations featuring a task to be learned by the system are observed and analysed by a Programming by Demonstration system developed at our institute in recent years. This section gives a short overview of the task acquisition process of our system. During the performance of a task in a so called training-center, the user is observed using diﬀerent sensors. These sensors include two data gloves for gathering the ﬁnger joint angles, magnetic trackers for obtaining the hands’ position, and two active stereo cameras mounted on pan-tilt-units for object localisation and tracking. Further details on the hardware used can be found in [6]. From the sensor input data, so called elementary operators are extracted. Elementary operators (EO’s) are actions that abstract a sensory motor coupling like primitive movement types (linear moves etc.), grasping and un-

54

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

Fig. 1. Hierarchical representation of manipulation segments

grasping actions. These EO’s abstract from the speciﬁc execution hardware and are robot-independent from a system point of view, although they have to be implemented on the speciﬁc target robot system. From the elementary operators, compound actions are constructed in a bottom-up manner. These are called macro-operators (MO’s) and contain the full task description. On a basis level, elementary move operators are aggregated to approach, depart and transport trajectories. Together with the grasp and ungrasp EO’s, grabbing and releasing of objects can be constructed. Between those gripper operations, various basic manipulation operations are detected (like transport operations, pouring operations and tool handling for instance). On the highest level, a sequence of manipulation segments is subsumed under a whole task demonstration (see ﬁgure 1). For further details, please refer to [16]. On each level in this hierarchical task representation, the changes to the environment are induced from lower levels of the hierarchy. The pre- and postconditions describing the environmental states before and after the execution of each macro-operator are propagated from the EO-level to the manipulation segment level and are computed from the environmental changes during the manipulation. The state of the environment is described in terms of ﬁrst order predicate logics, based on a set of semantical variables (see ﬁgure 2). These are basically geometrical inter-object relations (like “on”, “under”, “to the right of...” or “contained in ...”) and intra-object predicates like object class (“fork”, “plate”, “table”) and internal state descriptors (“opening angle”, “oven temperature”, “liquid level”, etc., depending on the object class). Each of these semantical variables is computed at every segmentation point, that starts or end a macro-operator. When the semantic variables change during the demonstration of a task these changes are accumulated to the overall eﬀects of the demonstration. From these eﬀects the pre- and postconditions of each node in the macro-operator tree, including the root, is computed. These are the central features of a task, used for the re-identiﬁcation of tasks and subtasks to be described in section 5.

Incremental Learning of Task Sequences

55

Fig. 2. Some semantical variables describing the relative positions of objects.

The system covers a broad range of manipulation classes drawn from the operations needed in a household environment: Transport operations (Pick&Place), device handling (operating household devices, opening a fridge) and tool handling (pouring liquids from a bottle to a glas, unscrewing a bottle etc.). It depends on the condition that the user is doing all changes in the environment (closed world assumption). The result of the task acquisition process outlined in this section is a task description, containing a sequence of manipulation segments together with their pre- and post-conditions and the hierarchical decomposition into elementary operators.

4 Learning Task Precedence Graphs This section is concerned with the representation of the sequential features of a task, the task precedence graph (TPG), and how it can be learned incrementally from multiple demonstrations of this task.

56

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

When a user performs a task, he performs it’s subparts (the manipulation segments) in a sequential ordering that he has chosen by random or by intent from all possible task execution orders. For a system recording his actions these appear as a simple sequence of operations. A learning system that builds knowledge about a task has to hypothesize on the underlying sequential task structure. These hypotheses can be represented by task precedence graphs, introduced by [11]. Deﬁnition 1. A task precedence graph (TPG) for a task T is a directed graph P = (N, R) with N being the set of subtasks o1 , o2 , . . . , on , and R ⊂ N × N being the set of precedence relations a faultless task execution must comply with. A precedence relation (o1 , o2 ) ∈ R with o1 , o2 ∈ N implies that the operation o1 must be complete before the execution of o2 can start. This is abbreviated as o1 → o2 . For a system that is supplied with only a single user demonstration D = (o1 , o2 , . . . , on ) of a task, it is hard to guess the inherent task precedence graph. The learning system can state diﬀerent equally valid hypotheses, ranging from the most restricting precedence graph P D = (N, RD ) that only allows the execution order demonstrated, to the most liberal TPG that posesses no sequential dependencies at all and allows the robot to choose the sequence of operations without any constraints. [11] states that the set of valid hypotheses is a version space, partially ordered by the subset-predicate on the sets of precedence relations R. While the learning system can not decide, which task precedence graph from the set of consistent TPG’s ﬁts the task best after seeing only a single example, it seems a viable approach to supply it with more sample demonstrations, applying diﬀerent task execution orders. In order to learn task knowledge from even a single demonstration suﬃcient for execution but improving the learned task when more knowledge in form of task demonstrations is available, an incremental approach is chosen. Assuming that, after processing m demonstrations {D1 , D2 , . . . , Dm } of the same task, the system has learned the most restrictive TPG Pm . Under observation of a new demonstration Dm+1 , the learned task precedence graph can be adapted in a way that incorporates the new knowledge. [8] suggests that this can be done by further generalising the previous hypothesis to a new one, covering the additional example. In order not to generalize too far, the minimal generalization of Pm that is consistent with Dm+1 must be chosen. The resulting new set of task precedence relations Rm+1 can be expressed as a function of the previously learned hypothesis Pm = (N, Rm ) and the most restrictive hypothesis for the new observed demonstration P Dm+1 = (N, RDm+1 ): (1) Rm+1 = Rm ∩ RDm+1 Now, one can state the process of incremental learning of task precedence graphs:

Incremental Learning of Task Sequences

57

1. For the ﬁrst training example D1 , initialize the task precedence graph to the most speciﬁc TPG P1 = P D1 . 2. For each additional demonstration Dm+1 of the same task, compute P Dm+1 and update the hypothesis Pm+1 according to equation 1. One issue not adressed until now is the question of when the additional task demonstrations arrive. In the application domain of household service robots one cannot assume that the user gives all demonstrations suﬃcient to learn the correct task precedence graph at once. Instead, it might take a long time between two successive demonstrations of the same task. Moreover, the task knowledge learned during past demonstrations has to be utilised because it is likely that the user will require the robot to execute the learned task without having taken into account all task demonstrations that might appear in the future. Therefore, the task precedence graph learned by the system should be reliable enough to ensure a faultless and, above all, secure task execution. With the mechanisms presented in this section it is possible to incrementally learn task precedence graphs from user demonstrations. The process of reﬁning task precedence graphs starts with the most restrictive hypothesis (= the order seen in the ﬁrst demonstrations) and continues to become more and more general as more demonstrations with diﬀerent sequences of actions are taken into account.

5 Subtask Similarity Measures Based on Information-Theoretic Metrics While the last section presented a method for learning the sequential constraints of a task when in every task demonstration exactly the same subtasks are used, this is not likely to be true for every task. Usually the human demonstrator will only use similar but partly diﬀerent manipulation segments in order to fulﬁll the same task. This section deals with the topic of identifying the corresponding manipulation segments in two or more demonstrations of the same task. In order to recognize the matching manipulation segments for two diﬀerent task demonstrations, the ordering of the manipulations can not be a reliable measure for subtask correspondence as the sequence of subtasks performed is potentially permuted. Though matching the segments that manipulate the same class of objects seems to be a good idea at ﬁrst, this method fails as soon as there are multiple objects of the same class present in the scene. So more features should be taken into account to determine the similarity of subtasks and establish suﬃciently robust subtask corespondences in order to identify operations of equal impact to the scene. In order to recognize the matching manipulation segments, two sets of features are taken into account:

58

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

• Object features: These contain the class of the objects manipulated or referenced in the certain subtask (cup, plate, table etc.) as well as their possible functional roles (liquid container, object container etc.). • Pre- and Postcondition features: These contain the geometrical relations that exist between the objects before or after the performance of the subtask respectively. The base operation to assess feature conformance is the measure of similarity sF for two sets of features A, B with the portion of common features in both feature sets: |A ∩ B| sF (A, B) = . (2) |A ∪ B| It turned out that this straight-forward approach exposes one major drawback: As the number of features is relatively high and many of the features are irrelevant to the certain task to be learned, the features carrying the relevant information will be predominated by the irrelevant ones. The solution is to weight each feature f depending of it’s relevance to the task to be learned with a weight w(f ). This turns equation 2 into sF (A, B) =

f ∈A∩B

w(f )

f ∈A∪B

w(f )

.

(3)

When choosing a useful weight function w(f ) one has to consider two points: Features of high relevance to the speciﬁc task should gain high weights. One way to assess the relevance of a feature is to test its occurrence over several diﬀerent demonstrations of the same task. Features with a high relevance to the task to be performed have a great probability of occurrence, while features of lower relevance will occurr less frequently. According to Shannon, the information content I(f ) of a feature f with probability p(f ) is I(f ) = − log2 p(f ). As features with low information content (= high probability of occurrence) to the speciﬁc task class T should be favored, w(f ) = − log2 (1 − p(f |T)) seems a reasonable choice for the weight function. On the other hand, when a completely new task is learned or only few demonstrations of the same task are available one has to take into account that only little information about the distribution of features in the speciﬁc task class is known. The idea is to introduce global background knowledge on the feature distribution. When several demonstrations of diﬀerent tasks have been observed by the robot and incorporated into the task knowledge base, it can be assumed that the features that uniquely discriminate the task are those that have low occurrence rates accross the whole task knowledge base. Thus, the information content to all other tasks T of feature f is − log2 p(f |T). An incremental learning system should be able to cope with both situations. Therefore a combination has to be found that favors the global information content measure when no or few task demonstrations are available, and the task-speciﬁc relevance measure as more demonstrations of the speciﬁc task become available. The weight function that fulﬁlls those requirements is

Incremental Learning of Task Sequences

w(f ) = − α log2 (1 − p(f |T)) + (1 − α) log2 p(f |T)

59

(4)

with the relative weighting α of the task-speciﬁc knowledge as α = 1−e−k·|T| . In our experiments we chose k = − ln 0.25 5 , such that after observing 5 demonstrations of the same task, the proportion α of task-speciﬁc to prior-knowledge is 0.75 : 0.25 and converging towards 1 for more demonstrations. This choice is motivated by [1], stating that the important features of a task can be sufﬁciently learned with 4 − 5 demonstrations. With the weighted proportion of features sF as in eq. 3 it is possible to compute the similarities spre and spost of the pre- and postcondition sets of two subtasks M1 and M2 and the average of the subtask similarities ssub of the tasks. They are set to spre = sF (P recond(M1 ), P recond(M2 )) spost = sF (P ostcond(M1 ), P ostcond(M2 )) 1 sM (m1 , m2 ). ssub = |M1 ∪ M2 | m1 ∈M1 ,m2 ∈M2

The (recursive) overall similarity sM of the two manipulation segments M1 , M2 is deﬁned as the simple average of the pre- and postcondition similarity and the average of all subtask similarities: sM (M1 , M2 ) = 13 · (spre + spost + ssub ). Once the subtask-similarity sM is computed for every pairing of the task’s subtasks, the subtask correspondence pST is computed as the permutation that maximizes the sum of similarities: pST = arg

sM (M1 , M2 ).

max

p∈perm(subtasks)

(M1 ,M2 )∈p

With the subtask permutation pST the most restrictive hypothesis P Di on the underlying task precedence structure can easily be constructed. This hypothesis can then be utilised to learn sequential task constraints in the way described in section 4.

6 Experimental Results In this section the experiments validating the method for incremental acquisition of task knowledge using heuristic relevance estimates for task features are reported. The task to be learned was chosen from the household domain; it consisted of laying the table for a main dish. The manipulated objects were: a plate, a fork and a mug. The main features that should be learned by the system are the spatial arrangement of the objects (the fork should be placed to the right of the plate and the mug placed on the distant side of the fork) as well as the sequential independence of all subtasks.

60

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

Eight diﬀerent demonstrations of that task were given (T = {D1 , D2 , · · · , D8 }). In the ﬁrst four of them, the fork, the plate and the mug were manipulated in this order. In the ﬁfth and sixth demonstration the mug was placed ﬁrst, followed by the fork and the plate. In the ﬁnal two demonstrations the mug, the plate and the fork were manipulated in this ordering. After taking into account the ﬁrst example, the initial task precedence graph P D1 = P1 was correctly obtained, which can be seen in the ﬁrst column of table 1. After observing a second example, D2 , the relevance estimates for the task similarity measure were continously taken into account. The relevance estimates for several features depending on the number of task demonstrations are depicted in ﬁgure 3(a).

0,9 0,8 0,7

0,025

0,6

0,02

Similarity

Relative Relevance Estimate

0,03

feat1 & 2 feat3 & 4 feat5 feat6 feat7

0,015 0,01

heuristical Spre naive Spre heuristical Ssub naive Ssub

0,5 0,4 0,3 0,2

0,005

0,1 0

0 1

2

3

4

5

6

7

1

8

2

3

4

5

6

7

8

#Task Demonstrations

# Task Demonstrations

(a)

(b) 1 0,9 0,8

Similarity

0,7 0,6

heuristical Spost naive Spost heuristical Sm naive Sm

0,5 0,4 0,3 0,2 0,1 0 1

2

3

4

5

6

7

8

# Task Demonstrations

(c) Fig. 3. (a) Relevance estimates for several features, depending on the number of tasks demonstrations taken into account. (b) & (c) resp.: Precondition-, Subtask-, Postcondition- and Overall-Similarity for the demonstrations D1 and D5 , calculated according to formula 3 (heuristic similarity) and 2 (naive similarity)

One can see that features 1-4 (the Relations mug ON table, fork ALIGNED EAST mug, plate ON table, fork ON plate) increase their relevance estimate as more task demonstrations become available. Meanwhile features 5-7 (the Relations fork TOUCHES SOUTH plate, mug TOUCHES NORTH

Incremental Learning of Task Sequences

61

plate and mug INTERSECTING fork ), occurring unintentionally in some task demonstrations, receive lesser ratings as they tend to occurr less frequently in all task demonstrations. With the assessment of feature relevance, the pre- and post-condition similarity spre resp. spost are calculated as well as the overall task similarity measure for all subtasks ssub , and ﬁnally the overall task similarity measure sM . In ﬁgures 3(b) & (c) , the results for the comparision for the task demonstrations D1 and D5 are shown in dependence of the number of task demonstrations regarded for the heuristical relevance estimation. The ﬁgures suggest the following: First, the heuristical measures start with a far better similarity than the naive approach. This is due to the second term in formula 4, which favours features that discriminate the new task from all other previously seen tasks. Second, the similarity measure gains higher values with more task-speciﬁc data becoming available. With more demostrations of the same task, the increasing weight α favors the ﬁrst part of equation 4. As the task feature statistics also become more striking with more demonstrations, this part ensures a task-speciﬁc adaption of the relevance measures. After the system has seen four examples with all the same order of actions, the learned task precedence graph is unchanged. After that, the system observes a demonstration featuring a diﬀerent execution order. The two different sequences and the recognized subtask permutation are shown in ﬁgure 4. The task precedence graph learned after ﬁve demonstrations is shown in the second column of table 1. When the last two user demonstrations are observed, the system can eliminate the last unnecessary precedence relation between the fork and the plate and is free to schedule every subtask as it is most convenient during the execution.

Fig. 4. Recognition of equal subtasks. Subtask correspondences are drawn with stipled lines. The sequential odering of subtasks is represented by bold arrows.

7 Conclusion This paper presents an approach to combining learning and reasoning on tasks. Tasks are recorded from user demonstrations, segmented, interpreted

62

M. Pardowitz, R. Z¨ ollner, and R. Dillmann

Table 1. Task precedence graphs learned incrementally during the performance of the experiment described in section 6. i

1-4

5-6

7-8

P Di

Pi

and stored in a data representation called macro-operators. Reasoning methods are applied to discover the task’s underlying sequential structure and reordering possibilities. Methods based on information-theoretic approaches are presented that estimate the relevance of task features. These heuristics combine pre-existing background knowledge on feature distributions across diﬀerent tasks with learned knowledge about the task itself. The methods and models outlined are demonstrated and validated in the context of an everyday houshold task to be learned by a robot servant. The results proved the utility of the relevance estimates, the task similarity measures and the algorithms for learning sequential constraints and opportunities of tasks. This approach presents a step towards robots that learn task knowledge from human demonstrations in an incremental manner that allow them to reﬁne and complete their learned knowledge as additional data becomes available, and act as real ’intelligent’ robot servants in future household applications.

References 1. Aude Billard, Yann Epars, Sylvain Calinon, Stefan Schaal, and Gordon Cheng. Discovering optimal imitation strategies. Robotics and Autonomous Systems, 47(2-3):69–77, 2004. 2. Cynthia Breazeal, Daphna Buchsbaum, Jesse Gray, David Gatenby, and Bruce Blumberg. Learning from and about others: Towards using imitation to bootstrap the social understanding of others by robots. Artiﬁcial Life, 11(1):31–62, 2005. 3. S. Calinon and A. Billard. Learning of Gestures by Imitation in a Humanoid Robot, chapter To appear, page In Press. Cambridge University Press, 2005. 4. Jason Chen and Alex Zelinsky. Programming by demonstration: Coping with suboptimal teaching actions. The International Journal of Robots Research, 22(5):299–319, May 2003. 5. R. Dillmann, O. Rogalla, M. Ehrenmann, R. Z¨ ollner, and M. Bordegoni. Learning Robot Behaviour and Skills based on Human Demonstration and Advice:

Incremental Learning of Task Sequences

6. 7. 8. 9.

10.

11. 12. 13. 14. 15.

16.

63

the Machine Learning Paradigm. In 9th International Symposium of Robotics Research (ISRR 1999), pages 229–238, Snowbird, Utah, USA, 9.-12. Oktober 1999. M. Ehrenmann, R. Z¨ ollner, O. Rogalla, S. Vacek, and R. Dillmann. Observation in programming by demonstration: Training and exection environment. In Humanoids 2003, Karlsruhe/Munich, Germany, October 2003, 2003. Y. Kuniyoshi, M. Inaba, and H. Inoue. Learning by Watching: Extracting Reusable Task Knowledge from Visual Observation of Human Performance. IEEE Transactions on Robotics and Automation, 10(6):799–822, 1994. Tom M. Mitchell. Generalization as search. Artiﬁcial Intelligence, 18(2):203– 226, 1982. M.. Nicoluscu and M. Mataric. Natural methods for robot task learning: instructive demonstrations, generalization and practice. In 2nd International joint conference on Autonomous agents and multiagent systems, 2003, pages 241–248, Melbourne, Australia, July 2003. H. Onda, H. Hirukawa, F. Tomita, T. Suehiro, and K. Takase. Assembly Motion Teaching System using Position/Force Simulator—Generating Control Program. In 10th IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 389–396, Grenoble, Frankreich, 7.-11. September 1997. M. Pardowitz, R. Z¨ ollner, and R. Dillmann. Learning sequential constratins of tasks from user demonstrations. In IEEE-RAS Intl. Conf. on Humanoid Robots (HUMANOIDS), Tsukuba, Japan, 2005. O. Rogalla. Abbildung von Benutzerdemonstrationen auf variable Roboterkonﬁgurationen. PhD thesis, University Karlsruhe, Departmen of Computer Science, 2002. Y. Sato, K. Bernardin, H. Kimura, and K. Ikeuchi. Task analysis based on observing handds and objects by vision. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, Lausanne, pages 1208 – 1213, 2002. S. Schaal. Is imitation learning the route to humanoid robots? In Trends in Cognitive Scienes, volume 3, pages 323–242, 1999. Qin Wang, Ruqing Yang, and Weijun Zhang. Skill acquisition from human demonstrations using fcm clustering of qualitative contact states. In Proc. 2003 IEEE Intl. Symp. on Computational Intelligence in Robotics and Automation, pages 937–942, 2003. R. Zoellner, M. Pardowitz, S. Knoop, and R. Dillmann. Towards cognitive robots: Building hierarchical task representations of manipulations from human demonstration. Intl. Conf. on Robotics and Automation (ICRA) 2005, Barcelona, Spain, 2005.

Representation, Recognition and Generation of Actions in the Context of Imitation Learning Haris Dindo1 and Ignazio Infantino2 1 2

DINFO, Dipartimento Ingengeria Informatica, Universit` a degli Studi di Palermo, Viale delle Scienze Ed. 6,90128, Palermo, Italy, [email protected] ICAR, Istituto di Calcolo e Reti ad Alte Prestazioni, Consiglio Nazionale delle Ricerche, Viale delle Scienze Ed. 11, 90128 Palermo, Italy, [email protected]

The paper deals with the development of a cognitive architecture for learning by imitation in which a rich conceptual representation of the observed actions is built. We adopt the paradigm of conceptual spaces, in which static and dynamic entities are employed to eﬃciently organize perceptual data, to recognize positional relations, to learn movements from human demonstration and to generate complex actions by combining and sequencing simpler ones. The aim is to have a robotic system able to eﬀectively learn by imitation and which has the capabilities of deeply understanding the perceived actions to be imitated. Experimentation has been performed on a robotic system composed of a PUMA 200 industrial manipulator and an anthropomorphic robotic hand.

1 Introduction Among the modern robotic control architectures, imitation systems have gained an important role, aiming to reduce the complexity of the robot programming by endowing the system with the capability to understand actions of a human demonstrator. For a review on learning by imitation paradigm and diﬀerent aspects on imitation see [15]. The learning by imitation approach involves several diﬃcult research problems as outlined in [4], [7] and [14]. In the past years, diﬀerent aspects of the problem were addressed in many working systems ([1], [2], [3], [10], [12], [17], [18]). However, the main emphasis is given to the problem of how to map the observed movement of a teacher onto the movement apparatus of the robot. While this is a fundamental aspect of any robotic system which copes with imitation, we claim that in order to eﬀectively learn by imitation, the system itself should have the capabilities to eﬀectively ”understand ” its environment and the perceived actions to be imitated. In our approach, understanding involves the generation of a high-level, declarative description of the perceived world. In other words, understanding

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 65–77, 2006. © Springer-Verlag Berlin Heidelberg 2006

66

H. Dindo and I. Infantino

copes with building an internal model of the external world. It endows the robot with the capability to perform reasonings and predictions about what surrounds it. As it has been proposed by Chella et al. [6], the conceptual spaces (see [8] for an introduction) allow to involve the generation of a high-level, declarative description of the perceived world. Following this approach, in this paper we propose a robust architecture that allows us to extract useful information from each intermediate level of a conceptual space for the purposes of imitation learning (geometric and color properties of objects, spatial relations between objects and actions performed on diﬀerent types of objects). The Moreover, learning and imitation are treated in the same architecture as two diﬀerent points of view of the same knowledge representation. The rest of the paper is organized as follows: in the next Section the cognitive architecture is summarized and detailed description of the conceptual representation is given. Section 3 shows an implementation of the architecture on a robotic platform. Finally, Section 4 gives short conclusions and outlines the future work.

2 The Cognitive Architecture The proposed architecture is organized in three computational areas as shown in ﬁg. 1. The Subconceptual Area is concerned with the low-level processing of perceptual data coming from vision sensors through a set of built-in visual mechanisms (e.g. color segmentation, feature extraction, teacher’s posture reconstruction), as well as with the control of the robotic system. In the Conceptual Area, diﬀerent properties of the perceived scene are captured: the Perceptual Space describes static properties of the entities in the scene (e.g. geometrical properties of the objects), the Situation Space encodes relations between various entities in a Perception (e.g. near, left, right, etc.), while the Action Space encodes the dynamic properties occurring in the world. In the Linguistic Area, representation and processing are based on a high-level symbolic language. The symbols in this area are anchored to sensory data by mapping them on appropriate representations in the diﬀerent layers of the conceptual area. The same architecture is used both to analyze the observed task to be imitated (observation mode) and to perform the imitation of the given task (imitation mode). For the purposes of the present discussion, we consider a simpliﬁed bidimensional world populated with various objects in which observation/imitation takes place. During the observation phase, the user shows her/his hand while performing arbitrary tasks of manipulating objects in front of a single calibrated camera. The task is then segmented into meaningful units and its properties (objects’ position and orientation, color and shape, relations between objects and action types) are represented, through the Conceptual Area, into the linguistic area in terms of high-level symbolic terms. In the

Representation, Recognition and Generation of Actions

67

Fig. 1. The computational areas of the proposed architecture and their relation with the external world.

imitation phase, the robotic system performs the same tasks triggered by a known stimuli in the external world or by a user’s query. In the following, a detailed description of each computational area is given. 2.1 The Subconceptual Area In the proposed architecture, the sub-symbolic processing area is organized in two distinct modules performing low-level processing of sensory data. The Perceptuomotor Module continuously tracks the actions of human demonstrator in the scene during the observation mode and reports any signiﬁcant changes in its motion state. During the imitation mode it translates high-level descriptions of the task to be performed into low-level control commands to be sent to the robot’s actuators. The Scene-Reconstruction Module detects all the objects in the scene and reconstructs them by means of parametric functions. It outputs positional, color and shape properties of objects in the scene. The Perceptuomotor Module The perceptuomotor module in the Subconceptual Area is concerned with both the human hand’s posture reconstruction and with the mapping between human and robot movements. The system performs skin detection in order to detect the human hand and continuously tracks its position and orientation in the image plane. The homography relationship between the image and the working plane of the user is computed during an oﬀ-line calibration phase. The module outputs the stimuli relative to the human hand’s trajectory in the world coordinate system:

68

H. Dindo and I. Infantino

P M Mhand (t) = [xhand (t), yhand (t), θhand (t)]T .

(1)

As the scene evolves changes occurs in the world because objects are being

(a)

(b)

(c)

(d) X Trajectory

300 250 200 150 100 50 0

0

1

2

3

4

5

6

4

5

6

Y Trajectory 300 250 200 150 100 50 0

0

1

2

3

Fig. 2. The upper ﬁgure shows the sequence of actions performed by the human user. Human hand is continuously tracked and its position and orientation is recorded. Zero-velocity crossing algorithm performs the detection of key-frames and objects’ properties (position, color and shape). The lower ﬁgure depicts computed trajectories of the human hand during the observation phase (cartesian position of the hand’s centroid). Vertical lines correspond to extracted key-frames.

grasped by the hand and moved to another location. In order to detect changes in a dynamic scene, we perform motion segmentation on the human hand data by using zero-velocity crossing detection. The zero-velocity occurs either when the direction of movement changes or when no motion is present. In order to accurately detect the zero-crossings in the movement data, we ﬁrst apply a low-pass ﬁlter to eliminate high-frequency noise components, and then we

Representation, Recognition and Generation of Actions

69

seek for the zero velocities within a threshold. The result is a list of frame indexes, named key-frames, in which object properties (shape and color) are to be processed. Fig. 2 shows a typical task performed by the user and the process of extracting the key-frames from the human hand trajectory data. This module is also responsible for the of the issues related to the mapping of teacher’s internal/external space, e.g. joint angles or Cartesian coordinates of the end-eﬀector, to that of the robot. In the current implementation we convert the Cartesian coordinates (in terms of x-y position and orientation) into robot joint angles by the means of standard inverse kinematics methods ([16]). Scene-Reconstruction Module For the purpose of the present discussion, we represent the objects in a scene as superellipses [11]. A superellipse is a closed curve deﬁned by the following equation: y 2 x 2 (2) ( )n + ( )n = 1 a b where a and b are the size (positive real number) of the major and minor axes and n is a rational number which controls the shape of the of the curve. For example, if n = 1 and a = b, we get the equation of a circle. For smaller n we get gradually more rectangular shapes, until for n → 0 the curve takes up a rectangular shape (ﬁg. 3). A superellipse in general position requires three additional parameters for expressing the rotation and translation relative to the center of the world coordinate system (x, y, θ). Therefore, in order to completely describe a superellipse, we need six parameters [a, b, n, x, y, θ]. For each detected object in the scene, we ﬁt a superellipse to it. First, we obtain the position and orientation of the objects in the scene by using the Hotelling transform (see [9]), and then we obtain the minimum bounding rectangle (i.e., the rectangle with minimum area that contains all the data). At this point, we are given all the superellipse parameters (axis lengths, center, and orientation) except the shape parameter, n, which can be approximated as ∼1−t± n=

t2 − 6t + 5,

(3)

Fig. 3. Superellipses with diﬀerent values of exponent n. Size parameters a and b are kept constant.

70

H. Dindo and I. Infantino

A where, t = 4ab , and A is the area of the image object(see [13]). An example of the ﬁtting process is given in the ﬁg. 4. Segmented Image (HSV Color Space Segmentation)

Fitted Superellipses

X=281.1;Y=180.0;Θ=−1.5° ID=3;GRASPED=0

X=105.0;Y=89.0;Θ=15.8° ID=1;GRASPED=0

X=247.0;Y=81.1;Θ=−19.7° ID=2;GRASPED=0

Fig. 4. Results of the superellipse ﬁtting algorithm.

However, since the process of superellipse ﬁtting is time-consuming, the detection of scene objects is performed only at the key-frames extracted in the Perceptuomotor Module. 2.2 The Conceptual Area The conceptual spaces are based on geometric treatment of concepts in which concepts are not independent of each other but structured into domains [8], e.g. color, shape, kinematics and dynamics quantities, and so on. Each domain is deﬁned in terms of a set of quality dimensions (e.g. the color domain may be described in terms of hue/saturation/value quantities while the kinematics domain is described by joint angles). In a conceptual space, objects and their temporal and spatial evolution are represented as points describing the quality dimensions of each domain. It is possible to measure the similarity between diﬀerent individuals in the space by introducing suitable metrics. In the

Representation, Recognition and Generation of Actions

71

context of present architecture we propose three conceptual spaces capturing diﬀerent temporal and spatial aspects of the scene (see ﬁg. 5) as described in the following.

Fig. 5. The structure of the Conceptual Area: Perceptual Space PS describes static properties of the entities in the scene (e.g. geometrical properties of the objects), the Situation Space SS encodes relations between various entities in a Perception (e.g. near, left, right, etc.), while the Action Space AS encodes the dynamical properties of the world.

The Perceptual Space Points in the Perceptual Space are computed starting from the stimuli obtained during the subconceptual processing stage. In this work, the quality dimensions of our Perceptual Space are given by the shape and color properties of the objects perceived: P Sobj = [aobj , bobj , mobj , hobj , sobj , vobj ]T .

(4)

Hence, this space captures the shape and color properties of objects. This is useful in assigning high-level terms to objects, such us square, circle, red, green, etc. Each point is the Perceptual Space is assigned a unique identiﬁer, ps id, together with the spatial information (x, y, θ) and the key-frame at which it has been recorded. The Situation Space The Situation Space encodes spatial relations between objects in a scene, such as left, right, near, etc. Suppose to have to have two objects in the Perceptual Space recorded at the same key-frame and described by their labels: (ps id1 , x1 , y1 , θ1 , kf and (ps id2 , x2 , y2 , θ2 , kf ). The situation between objects is then described by using the following equation: x = x1 − x2 ,

y = y1 − y 2 ,

θ = θ 1 − θ2 .

(5)

72

H. Dindo and I. Infantino

Hence, the quality dimensions of the space are given by the ( x, y, θ), and it results to be symmetric with respect to the origin. Each point in the Situation Space is assigned a unique identiﬁer, ss id, together with the ordered set of Perceptual Space labels {ps id1 , ps id2 }. A particular ss id is used to describe the absolute position/orientation of objects at each key-frame. It is straightforward to introduce suitable metrics in order to measure spatial relations between objects in a scene. The Action Space In order to account for the dynamic aspects of the perceived scene, we propose an Action Space in which each point represents the simple motion of entities. An action corresponds to a ”scattering” from one situation to another in the conceptual space. However, the decision of which kind of motion can be considered simple is not straightforward, and it is strictly related to the problem of motion segmentation, as described in Sect. 2.1. An immediate choice is to consider ”simple” any motion of the human hand occurring between keyframes. In our world there are two possible types of actions which can be performed by the hand: reach (reaches an object for the purposes of grasping it) and transfer (moves the hand to an absolute position and orientation in the world coordinate system, or relative to other objects in the world). In the ﬁrst case, the trajectory of the hand is expressed relative to the object to be grasped, while in the second, it is represented relative to the origin of the world reference system. In both cases, we will refer to target conﬁguration, xtarget . In order to represent simple motions, it is possible to choose a set of basis functions, in terms of which any simple motion can be expressed. Such functions can be associated to the axes of the dynamic conceptual space as its dimensions. Therefore, from the mathematical point of view, the resulting Action Space is a functional space. A single point in the space therefore describes a simple motion of the human hand. In the context of the present work we choose the technique of Principal Component Analysis (PCA) in order to obtain the set of basis functions. We ﬁrst obtain a set of training trajectories performed by diﬀerent users, ϕ(t) = [x(t), y(t), θ(t)]T , and represent each trajectory relative to the target conﬁguration, ϕ (t) = ϕ(t) − xtarget . In order to account for diﬀerent temporal durations of simple motions, we interpolate the data to a ﬁxed length. We noted that generally the ﬁrst six principal components account for the most of the data variance, which are used to encode the simple human hand motions. Principal components form the basis functions, ϕi (t), in terms of which each hand trajectory can be represented as a linear combination: k

ci ϕi (t)

ϕ (t) = i=1

(6)

Representation, Recognition and Generation of Actions

73

where ci are the projection coeﬃcients, and the index k refers to the number of principal components (in our case, k = 6). Hence, the quality dimensions of the Action Space are given by the PCA coeﬃcients: AS = [c1 , c2 , . . . , c6 ]T .

(7)

However, principal components are useful also as the basis for generating novel trajectories that resembles those seen during the training phase. This is done as the process of data interpolation: given the starting, (ϕ (t0 )), and ending, (ϕ (tf )), hand conﬁgurations, it is possible to determine the coeﬃcients [c1 , . . . , c6 ]T which satisfy the boundary conditions by solving the set of six (linear) equations in six unknowns: k

k

ϕ (t0 ) =

ci ϕi (t)

ci ϕi (t); ϕ (tf ) = i=1

(8)

i=1

2.3 The Linguistic Area The more “abstract” forms of reasoning, that are less perceptually constrained, are likely to be performed mainly within the Linguistic Area. The elements of the Linguistic Area are symbolic terms that have the role of summarizing perceptions and actions represented in the conceptual spaces previously described, i.e., linguistic terms are anchored to the structures in the conceptual spaces [5]. We adopted a ﬁrst-order logic language in order to implement the Linguistic Area. Perceptions and actions are represented as ﬁrst-order terms. The semantics of the language is given by structures in conceptual spaces according to G¨ardenfors [8]. In particular, terms related to actions are interpreted in the action space, terms related to the spatial properties of the objects in the scene are interpreted in the situation space, while terms related with objects are interpreted in the perceptual space. In the present work, we adopt a ﬁnite number of primitive actions the robot can perform: • reach and grasp(ps id, as id), adopted to describe an action of reaching and grasping the object described by the identiﬁer ps id by executing the action as id, and • transf er and release(ss id, as id), adopted to describe an action, described by the indentiﬁer as id, of transferring and releasing a grasped object to an absolute position/orientation or relative to other objects in the workspace. Each complex task may be described as a sequence of primitive actions: task : −reach and grasp(ps id, as id), transf er and release(ss id, as id), ...

74

H. Dindo and I. Infantino

3 Experimental Results Our robotic setup is composed of a conventional six degrees-of-freedom PUMA 200 manipulator and an anthropomorphic four-ﬁngered robotic hand. The overall system has 22 degrees-of-freedom, and it is endowed with the ability to navigate in the free space, and to grasp objects which can be described as superellipses. The experiments we performed are concerned with the problem of teaching the robotic system several manipulative tasks. In order to facilitate the process of teaching the robot a task, we have developed the Visual Blackboard System for this purpose. It allows a user to create a workspace with the same dimensions as the real one, to populate it with various objects, and to execute several tasks. It also allows to observe the results of the imitation phase. The Visual Blackboard System is shown in ﬁgure 6.

Fig. 6. Visual Blackboard System for teaching tasks.

As an example, consider the task of placing the ”red square” to the right of the ”green circle”. Having seen this task performed by a user, the system extracts the key-frames and encodes all scene objects, their relations and human hand motion in the Conceptual Area. Suppose that ”red square” is assigned the perceptual space label ps id1 and green circle is assigned the label ps id2 . Their ﬁnal situation is assigned the situation space label ss idgoal .

Representation, Recognition and Generation of Actions

75

This task is automatically assigned the following description: reach(ps id1 ), grasp(ps id1 ), transfer relative to(ss idgoal ), and could be labelled accordingly to the goal performed. Since the perceptual space points hold the whole information about the objets involved in a task, it is possible to refer the previous description to another scene in which similar object are present (e.g. red square and green circle) and perform exactly the same task. However, the main strength of our architecture lies in the fact that it is also possible to anchor the same task to the more general instances by exploiting the conceptual representations. Hence, it is possible to obtain the task of ”placing squares left to circles”, or ”placing red objects left to the green ones”, and their combinations. However, each task description is given a priority: less general tasks have higher priority with respect to more general tasks. In the Imitation Mode, the system observes the scene and searches for an already seen task according to the priority (in our example it suﬃces to seek for squares, circles, red or green objects). If it is not able to match the task, it prompts the user which eventually shows the task in the observation mode. Otherwise, the matched description is then tailored to the particular situation, and absolute positions and orientations to be reached by the robotic system are produced and sent to the Action Space for the trajectory computation. The whole trajectory is then sent to the Robotic Control Module and mapped to our robot’s actuators. An example of the robot imitating the task (depicted in ﬁg. 2) is shown in ﬁg. 7. Note that in this example, the control of the robotic hand ﬁngers is performed as a separate behavior and is not involved in the imitation process.

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 7. (a) Top view of the world as seen by the robot. (b)-(f) Sequence of actions performed by the robot in the imitation mode: the robot executes the same task performed by the user.

76

H. Dindo and I. Infantino

4 Conclusions The main contribution of the paper is the proposal of an eﬀective cognitive architecture for imitation learning in which three diﬀerent representation and processing areas coexist and integrate in a theoretical way: the subconceptual, the conceptual and the linguistic areas. The resulting system is able to learn natural movements and to generate complex action plans by manipulating symbols in the linguistic area. It should be noted that the conceptual spaces may be employed as a simulation structure, i.e. to simulate movements and actions before eﬀectively executing them. Current research concerns how to enhance the proposed architecture by stressing this simulating capabilities in the context of imitation learning.

References 1. C. G. Atkeson, S. Schaal (1997) Learning Tasks From A Single Demonstration. In: Proceedings of IEEE-ICRA, pp. 1706-1712, Albuquerque, New Mexico. 2. A. Billard, M.J. Mataric (2001) Robotics and Autonomous System, no. 37, pp. 145-160. 3. A. Billard (2002) Imitation: a means to enhance learning of a synthetic protolanguage in autonomous robots. In: K. Dautenhahn and C.L. Nehaniv (eds), Imitation in Animals and Artifacts, pp. 281-310, MIT Press. 4. C. Breazeal, B. Scasselati (2002) Challenges in Building Robots That Imitate People. In: K. Dautenhahn and C.L. Nehaniv (eds), Imitation in Animals and Artifacts, pp. 363-390, MIT Press. 5. A. Chella, M. Frixione, S.Gaglio (2003) Robotics and Autonomous Systems, Special Issue on Perceptual Anchoring, vol. 43, 2-3, pp. 175-188. 6. A. Chella, H. Dindo, I. Infantino (2005) A Cognitive Framework for Learning by Imitation. ICRA05 workshop on the Social Mechanisms of Robot Programming by Demonstration, Barcelona, Spain. 7. K. Dautenhahn, C.L. Nehaniv (2002) The agent-based perspective on imitation. In: K. Dautenhahn and C.L. Nehaniv (eds), Imitation in Animals and Artifacts, pp. 1-40, MIT Press. 8. P. G¨ ardenfors (2000) Conceptual Spaces. MIT Press-Bradford Books, Cambridge, MA. 9. R. C. Gonzales, and P. Wintz (1987) Digital Image Processing (2nd ed.). Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA. 10. J.A. Ijspeert, J. Nakanishi, S. Schaal (2002) Movement imitation with nonlinear dynamical systems in humanoid robots. In: Proceedings of Intl. Conf. on Robotics and Automation, Washington. 11. A. Jaklic, A. Leonardis, F. Solina (2000) Segmentation and Recovery of Superquadrics. Computational Imaging and Vision, Kluwer, Dordrecth, 2000 12. Y. Kuniyoshi, M. Inaba, and H.Inoue (1994) IEEE Trans. Robot. Automat., 10:799-822, November. 13. P. Rosin (2000) IEEE Transactions. on Pattern Analysis and Machine Intelligence, Vol. 22, No. 7, pp.726-232. 14. S. Schaal (1999) Trends in Cognitive Sciences 3, pp.233-242.

Representation, Recognition and Generation of Actions

77

15. S. Schaal, A. J. Ijspeert, A. Billard (2003) Computational Approaches to Motor Learning by Imitation. In: Philosophical Transactions: Biological Sciences (The Royal Society), no. 358, pp.537-547. 16. L. Sciavicco, and B. Siciliano (2000) Modelling and Control of Robot Manipulators. 2nd ed. pringer-Verlag, New York. 17. A. Ude, T. Shibata, C.G. Atkeson (2001) Robotics and Autonomous Systems, vol. 37, pp. 115-126. 18. D. M. Wolpert, K. Doya, M. Kawato (2003) A unifying computational framework for motor control and social interaction. In: Philosophical Transactions: Biological Sciences (The Royal Society), no. 358, pp. 593-602.

Reduction of Learning Time for Robots Using Automatic State Abstraction Masoud Asadpour1 , Majid Nili Ahmadabadi2 , and Roland Siegwart1 1 2

Autonomous Systems Lab, Ecole Polytechnique F´ed´erale de Lausanne (EPFL), CH-1015 Lausanne, Switzerland {masoud.asadpour,roland.siegwart}@epfl.ch Control and Intelligent Processing Center of Excellence, ECE Dept., University of Tehran, Iran [email protected]

Summary. The required learning time and curse of dimensionality restrict the applicability of Reinforcement Learning(RL) on real robots. Diﬃculty in inclusion of initial knowledge and understanding the learned rules must be added to the mentioned problems. In this paper we address automatic state abstraction and creation of hierarchies in RL agent’s mind, as two major approaches for reducing the number of learning trials, simplifying inclusion of prior knowledge, and making the learned rules more abstract and understandable. We formalize automatic state abstraction and hierarchy creation as an optimization problem and derive a new algorithm that adapts decision tree learning techniques to state abstraction. The proof of performance is supported by strong evidences from simulation results in nondeterministic environments. Simulation results show encouraging enhancements in the required number of learning trials, agent’s performance, size of the learned trees, and computation time of the algorithm. Keywords: State Abstraction, Hierarchical Reinforcement Learning

1 Introduction It is well accepted that utilization of proper learning methods reduces robots’ after-design problems caused naturally due to use of inaccurate and incomplete models and inadequate performance measures at the design time in addition to confronting unseen situations and environment and robot changes. Existing learning methods that are capable of handling dynamics and complicated situations are slow and produce rules – mostly numerical – that are not abstract, modular, and comprehensible by human. Therefore, the designer’s intuition cannot be simply incorporated in the learning method and in the learned rules. In addition, the existing learning methods suﬀer from the curse of dimensionality, requiring a large number of learning trials and lack of abstraction capability. RL [1] despite its strength in handling dynamic and nondeterministic environments is an example of such learning methods. In this

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 79–92, 2006. © Springer-Verlag Berlin Heidelberg 2006

80

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

Fig. 1. State abstraction concept

research we try to solve some of the mentioned problems in RL domain because of its simplicity and variety of its applications in robotics ﬁeld, in addition to its mentioned advantages. It is believed that some of the discussed drawbacks can be lessen to a great extend by implementation of state abstraction methods and hierarchical architectures. Having abstract states and hierarchies in the agent’s mind, the learned rules are partitioned, deﬁnition of performance measures for future learning is simpliﬁed, and the number of learning trials can be reduced by a careful design of learning method. In addition, incremental improvement of agent’s performance becomes much simpler. Diﬀerent approaches have been proposed so far for state abstraction and creation of hierarchies and implementation of RL methods in those architectures. One group of these methods assumes a known structure and use RL methods to learn both rules in each layer and the hierarchy. Another group of researches proposes hand-design of subtasks and learning them in a deﬁned hierarchy. These two approaches require much design eﬀort for explicit formulation of hierarchy and state abstraction.Moreover, the designer should to some extend know how to solve the problem before s/he designs the hierarchy and subtasks. The third approach uses decision-tree learning methods to automatically abstract states, detect subtasks and speed up learning (Fig. 1). The devised approaches however do not take the exploratory nature of RL into account which results in non-optimal solution. In this paper, we take a mathematical approach to develop new criteria for utilization of decision trees in state abstraction. Our method outperforms the existing solutions in performance, number of learning trials, and size of trees. Related works are reviewed in the next section. U-Tree algorithm and its drawbacks are brieﬂy described in the third section. In the fourth section we formalize state abstraction as an optimization problem and derive new split criteria. The ﬁfth section describes the simulation results. Conclusions and future works are discussed ﬁnally.

Reduction of Learning Time for Robots Using Automatic State Abstraction

81

2 Related Works The simplest idea to gain abstract knowledge is detection of subtasks, specially repeating ones [2], and discovering the hierarchy among them. Once the agent learned a subtask, which is faster and more eﬃcient due to confronting with smaller state space, it would be able to use it afterwards. Subtasks in Hierarchical RL(HRL) are closed-loop policies that are generally deﬁned for a subset of the state set [3]. These partial policies are sometimes called temporallyextended-actions, options [4], skills [5], behaviors [6], etc. They must have well-deﬁned termination conditions. Although these methods give more understanding about the underlying aspects of Hierarchical RL however; since particular features are designed manually based on the task domain, it is hard to apply them to complex tasks. The general problem that should be solved ﬁrst, either when designing manually or extracting the structures automatically is the state abstraction. 3 In many situations signiﬁcant portions of a large state space may be irrelevant to a speciﬁc goal and can be aggregated into a few, relevant, states. As a consequence, the learning agent can learn the subtask reasonably fast. Our work addresses this problem. Techniques for non-uniform state-discretization are already known e.g. Parti-game [8], G algorithm [9], and U-Tree [10]. They start with the world as a single state and recursively split it when necessary. Continuous U-Tree [11] extends U-Tree to work with continuous state variables. TTree [12] applies Continuous U-Tree to SMDPs. Jansson and Barto [13] applied U-Tree to options but they use one tree per option and build the hierarchy of options manually. We adopt U-Tree as our basic method because the reported results illustrate it as a promising approach to automatic state abstraction. In this paper we show that the employed U-Tree methods ignore RL’s explorative nature. This imposes a bias in the distribution of samples that are saved for introducing new splits in U-Tree. As a result, ﬁnding a good split becomes more and more diﬃcult and the introduced splits can be far from optimal. Moreover, U-Tree-based techniques have been excerpted in essence from decision tree learning methods. Therefore the used split criteria are very general and can work with any sort of data. Here we devise some specialized split criteria for state abstraction in RL and show that they are more eﬃcient than the general ones both in learning performance and computation time. Dean and Robert [14] have developed a minimization model, known as reduction, to construct a partition space that has fewer number of states than the original MDP, while ensuring that the utility of the policy learned in the reduced state space is within a ﬁxed bound of the optimal policy. Our work is diﬀerent because -reduction does not extract any hierarchy among state 3

The inverse of state abstraction, State Aggregation [15] clusters “similar” states and assigns them same values. It eﬀectively reduces the state space size, but requires large number of trials and huge initial memory.

82

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

partitions, while building a hierarchy can reduce search time in partition lists from O(n) to O(log n), n being the number of partitions. Also, their theory is developed based on immediate return of actions instead of long-term return. We can argue that -reduction is a special case of our method.

3 Formalism We model the world as a MDP which is a 6-tuple (S, A, T, γ, D, R) , where S is a set of states (here can be inﬁnite), A = {a1 , . . . , an } is a set of actions, a T = {Ps´ s } is a transition model that maps S ×A×S into probabilities in [0, 1], γ ∈ [0, 1) is a discount factor , D is the initial-state distribution from which the start state is drawn (shown by s0 ∼ D ), and R is a reward function that maps S × A × S into real-valued rewards. A policy π maps from S to A, a realvalued value function V on states, or a real-valued action-value function (also called Q-function) on state-action pairs. The aim is to ﬁnd an optimal policy π∗ (or, equivalently, V ∗ or Q∗ ) that maximizes the expected discounted total rewards of the agent. We assume that each state s is a sensory-input vector (x1 , . . . , xn ), where xi is a state-variable (called also feature, or attribute). 3.1 U-Tree A U-Tree [10] [11] abstracts the state space incrementally. Each leaf Lj of the tree corresponds to an abstract state S¯j . Each leaf keeps action-values Q(S¯j , a) for all available actions. The tree is initialized with a single leaf, assuming the whole world as one abstract state. New abstract states are added if necessary. Sub-trees of the tree represent subtasks of the whole task. Each sub-tree can have other sub-sub-trees that correspond to its sub-sub-tasks. The hierarchy breaks down to leaves that specify primitive sub-tasks. Continuous U-Tree [11] loops through a two phase process: sampling and processing phases. In sampling phase, a history of the transition steps Ti = (si , si , ri , s´i ) composed of the current state, the selected action, the received immediate reward, and the next state is saved. The sample is assigned to a unique leaf of the tree, based on the value of the current state-variables. During the sampling phase the algorithm behaves as a standard RL algorithm, with the added step of using the tree to translate sensory input to an abstract state. After some number of learning episodes the processing phase starts. In processing phase a value is assigned to each sample: V (Tt ) = rt + γV (¯ st+1 ), V (¯ st+1 ) = max Q(¯ st+1 , a) a

(1)

where s¯t+1 is the abstract state that st+1 belongs to. Then if a signiﬁcant diﬀerence among the distributions of sample-values within a leaf is found that leaf is split in two leaves.

Reduction of Learning Time for Robots Using Automatic State Abstraction

83

To ﬁnd the best split point for each leaf, the algorithm loops over statevariables. The samples within a leaf are sorted according to a state-variable and a trial split is virtually added between each consecutive pair of statevariable values. This split divides the abstract state in two sets. A splitting criterion, like Kolmogorov-Smirnov test (KS), compares the two sets and returns a number indicating the diﬀerent between the two distributions. If the largest diﬀerence among all state-variables is bigger than a conﬁdence threshold, then a split is introduced. This procedure is repeated for all leaves. Although the original algorithm for U-Tree [10] can function in partially observable domains, for the sake of simpliﬁcation, we make the standard assumption that the agent has access to a complete description as well. Also, since the structure of the tree is not revised once a split is introduced; we assume the environment is stationary. If we assume processing phase is done after each learning episode, the number of leaves 4 at episode t is L(t), and sample size in each leaf is m, the best sort algorithms, e.g. Quicksort, can sort the samples based on a statevariable in O(m log m). Maximum split points can be (m − 1). To ﬁnd the cumulative distribution, KS-test needs the samples in the two virtual splits to m be sorted based on their values. This needs at least O(2 m 2 log 2 )). Then a pass through all sorted samples needs O(m). Totally the order of the algorithm using KS-test and assuming n being dimension of the state vectors, is: O(L(t) n m4 log m log

m ) 2

(2)

3.2 Splitting Criteria In literature two types of split criteria have been used for state abstraction. Rank-based tests e.g. non-parametric KS-test [10] [11], and tests inspired from Information Theory e.g. Information Gain Ratio(IGR) [17]. Non-parametric KS-test looks for violation of Markov Property by computing statistical difference between the distributions of samples on either side of a split using the diﬀerence in cumulative distributions of the two datasets. IGR-test measures the amount of reduction in uncertainty of sample values and is computed as follows; given a set of samples T = {Ti = (si , ai , ri , s´i )}, i = 1, . . . , m from ab¯ labeled with V (Ti ) – computed according to (1) – the Entropy stract state S, of sample-values is: EntropyS¯ (V ) = − v

¯ log2 P (v(Ti ) = v|S) ¯ P (V (Ti ) = v|S)

(3)

¯ is the probability that samples in T have value v. where P (V (Ti ) = v|S) Information Gain of splitting the abstract state S¯ to S¯1 and S¯2 is deﬁned as the amount of reduction in the entropy of sample-values i.e.: 4

Total number of nodes is 2L(t) − 1.

84

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

¯ S¯1 , S¯2 ) = EntropyS¯ (V ) − Inf oGain(S,

2

EntropyS¯i (V ) i=1

mi m

(4)

where mi is the number of samples from T that fall in S¯1 . IGR is deﬁned as: ¯ S¯1 , S¯2 ) = IGR(S,

¯ S¯1 , S¯2 ) Inf oGain(S, −

2 mi i=1 m

log2

mi m

(5)

3.3 Biased Sampling Problem Each leaf of U-Tree in implementation can hold a limited number of samples. If the number of samples exceeds an upper limit, an old sample is selected randomly and is replaced by the new sample. However, selection probability of an action in a state depends on its action-value function. This fact aﬀects the distribution of samples in a leaf. We call this a biased sampling. The biased sampling problem has a negative eﬀect; RL algorithms escape from local optimums by making a trade oﬀ between exploration and exploitation. If size of a sample list is not big enough, then number of samples with explorative character would not be statistically signiﬁcant to introduce a split, until the conﬁdence threshold for split criterion is reduced. Reducing the threshold increases the occurrence of non-optimal splits. One expect that size of the generated tree might be large with low performance. This problem can be solved by having in each leaf, instead of one big list with capacity k, |A| k smaller lists with capacity |A| per action. Then the algorithm is modiﬁed so that when new samples arrive and the sample list is full, an old sample from the list that corresponds to the action of the new sample is deleted randomly. Now sorting the new sample list would be faster: |A| O(

k k k log ) = O(k log ) < O(k log k) |A| |A| |A|

(6)

4 SANDS In this section we formalize the state abstraction problem and provide mathematical description of a new algorithm that is specially derived for HRL. We call it SANDS which stands for State Abstraction in Nondeterministic Systems. Also we show how we can combine SANDS with other heuristics. 4.1 State Abstraction An abstract state S¯ is a subset of state space S, so that all states within it have “close” values (see bellow). Value of abstract state S¯ is deﬁned as the expected discounted reward return if an agent starts from a state in S¯ and follows the policy π afterwards i.e.:

Reduction of Learning Time for Robots Using Automatic State Abstraction

¯ = V π (S)

¯ P (s|S) ¯ s∈S

a a π Ps´ s)) s (Rs´ s + γV (´

π(s, a) s∈A

85

(7)

s´

where π(s, a) is the selection probability of action a in state s under policy π, a Ps´ ´ after doing action a s is the probability that environment goes to state s a in state s, and Rs´ is the expected immediate reward after doing action a in s ¯ a) is deﬁned as: state s and going to state s´. Similarly, Qπ (S, ¯ a) = Qπ (S,

¯ P (s|S) ¯ s∈S

a a π Ps´ s)) s (Rs´ s + γV (´

(8)

s´

From set theory, a set partition of a set S, is deﬁned as a collection of disjoint subsets of S whose union is S. The number of partitions of the set is called a Bell number. So U = {S¯1 , . . . , S¯p } is a set partition of S iﬀ: S¯i ⊆ S, ∀i ∈ {1, . . . , p} p S = i=1 S¯i ¯ Sj = ∅, ∀i, j ∈ {1, . . . , p}, i = j

S¯i

(9)

State abstraction is then an optimization problem which aims to ﬁnd a set partition and a policy π on state space S with p, the Bell number of set partition, minimized while ensuring that the value of the learned policy π is within a ﬁxed bound of the optimal policy π∗ , i.e.: ∗ |V π (s) − V π (S¯i )| < , ∀s ∈ S¯i , ∀S¯i ∈ U

(10)

4.2 SANDS Splitting Criterion It is easy to prove that leaves of U-Tree generate a set-partition of state space. It is clear from deﬁnition of state abstraction in (10) that we are minimizing the diﬀerence between the optimal policy and the policy that U-Tree represents. Here we, instead of walking toward the optimal policy which is unknown, try to walk away from the current policy toward the optimal policy. Now, assume that we have a U-Tree that deﬁnes a policy π. We would like ˜ by splitting the partition S¯ (one of the leaves in to enhance it to a policy π the tree) to unknown partitions S¯1 and S¯2 , so that the expected return of the resulting tree is maximized i.e. we maximize: P (s)[V ˜π (s) − V π (s)]

G=

(11)

s∼D

where D is initial states distribution. Using Bellman equations we can write: G=

P (s) s∼D

−

˜ [π(s, a) a

a [π(s, a)

a a s s (Rs´ s´ Ps´ a a (R P s´ s s s´ s´

s))] + γV ˜π (´ + γV π (´ s))]

(12)

86

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

s∼D a

a ˜ Rs´ s (π(s, a) − π(s, a)) ˜(s, a)V ˜π (´ +γ(π s) − π(s, a)V π (´ s))

a P (s)Ps´ s

G= s´

(13)

V ˜π (s) is unknown, we can use V π (s) as an estimation. For the states that fall ¯ π ˜(s, a) = π(s, a). outside of S¯ policy remains unchanged after split i.e ∀s ∈ S, ¯ ¯ For the states within S or within the new abstract states S1 and S¯2 we deﬁne ˜i (a) = π ˜(s, a), s ∈ S¯i , i = 1, 2. Hence: π(a) = π(s, a), s ∈ S¯ and π   2

G= a

( π ˜i (a) − π(a)) ¯i s∈S

i=1

a a π P (s)Ps´ s)) s (Rs´ s + γV (´

(14)

s´

¯ × From the properties of conditional probabilities P (s) = P (s|S¯i ) × P (S¯i |S) ¯ ∀s ∈ S¯i , i = 1, 2. Considering the deﬁnition in (8), we can write: P (S), 2

¯ G = P (S) a

¯ ˜i (a) − π(a))Qπ (S¯i , a)P (S¯i |S) (π

(15)

i=1

This way the learner ﬁnds the split that maximizes the performance of the tree over the “whole action set”, provided that actions are selected according to Softmax distribution. But an important point here is that the learner can not revise the tree after initiating the split. This means that it is ﬁxing somehow a part of the policy in this level. Therefore, it is more valuable to judge the splits based on the performance of the greedy action instead of Softmax. The optimization term then should change from summation to maximum. More¯ is same for all states in abstract state S. ¯ So, the ﬁnal formulation over, P (S) for SANDS criterion is simpliﬁed to: ¯ S¯1 , S¯2 ) = max SAN DS(S, a

2

¯ ˜i (a) − π(a))Qπ (S¯i , a)P (S¯i |S) (π

(16)

i=1

U-Tree generates a set-partition of state space. But, optimality of the number of partitions is not guaranteed although the algorithm starts with minimum partitions and adds more partitions if required. In fact the splits here are Univariate [16] that shrink the state space parallel to one of its axes. The necessary (but not suﬃcient) condition to have minimum size tree is to have Multivariate [16] splits, which can split the space in any direction. Multivariate splits are not covered in the current version of the algorithm however; we discuss the size of the trees in result section. If we assume processing phase starts after one learning episode, number of leaves in episode t is L(t), sample size in each leaf is m, and state dimension is m n, time order of the new algorithm would be O(L(t) n m (m2 − |A|2 ) log |A| ) which is comparable to (2) by a rough factor of m log m. 4.3 Using Mont Carlo Method to Estimate SANDS We now show how to compute each term by Mont-Carlo method using the samples recorded during learning episodes. Each logged sample is a 4-tuple

Reduction of Learning Time for Robots Using Automatic State Abstraction

87

(si , ai , ri , s´i ), corresponding to current state, selected action, next state, and received reward. If among ma samples for action a in S¯ , ma1 and ma2 = ma −ma1 samples correspond to S¯1 and S¯2 respectively, we can compute: ρˆai = ˆ ai = µ ˆai = π

ma i ma , i 1 ma i

= 1, 2 ma i j=1 (rj

ρˆi = + γ maxb Q(´ sj , b)), i = 1, 2

ˆa exp(µ i /τ) P ,i ˆb b exp(µi /τ)

= 1, 2

ˆ a = ρˆa1 µ ˆ a2 ˆ a1 + ρˆa2 µ µ ˆa /τ) exp(µ P ˆb b exp(µ /τ) a a ˆ(a))µ ˆ i ρˆi π

ˆ(a) = π

= 1, 2

¯ S¯1 , S¯2 ) = maxa SAN DS(S,

P ma Pa ia , i am

2 ˆ i=1 (πi (a)

−

(17)

4.4 SoftMax Gain Ratio (SMGR) New criteria can be derived by combining SANDS with diﬀerent heuristics. One of them is what we call SoftMax Gain Ratio: 2 aˆ ˆ ˆ ˆ ¯ S¯1 , S¯2 ) = max −π(a) log π(a) + i=1 ρˆi πi (a) log πi (a) SM GR(S, 2 a − i=1 ρˆi log ρˆi

(18)

In this criterion, instead of deﬁning entropy function on sample-values, we deﬁne it on action-probabilities and try to ﬁnd the splits that maximize the reduction of this entropy i.e. the split that results in more certainty on selection probability of actions. We can prove that SMGR scales and combines SANDS with a penalty term as following: ¯ S¯1 , S¯2 ) = SM GR(S,

1 ¯ ¯ ¯ a τ SAN DS(S, S1 , S2 ) + log 2 − i=1 ρˆi log ρˆi

2 i=1

πρiˆi (a)

(19)

We can show the penalty term punishes the splits that result in: (1) giving high probability to a speciﬁc action without support of enough evidence (samples), and (2) contradicting estimations on selection probabilities of a speciﬁc action in either side of the virtual split, in the sense that an action is the best action in one split but the same action is the worst action in the other one.

5 Simulation Results We have tested our algorithms on a simpliﬁed football task. It is very similar to the classic Taxi problem with the possibility of having moving players with diﬀerent patterns to create environments with diﬀerent levels of nondeterministicity. The playing ﬁeld is an 8 × 6 grid (Fig. 2). The x axis is horizontal and y axis is vertical. There are two goals located at (0, 3) and (8, 3). A learning agent plays against two “passive” opponents and learns to deliver the ball to the left goal. By passive opponents we mean they move in the ﬁeld only to restrict movements of the learner and do not perform any action on the ball. The learner can choose among 6 actions: Left, Right, Up, Down, Pick,

88

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

Fig. 2. A screen-shot from the simulator

Put. If any player selects an action that results in touching the boundaries or going to a pre-occupied position by another player, that action is ignored. In a special case when the learner has picked the ball, if it hits the other players or a boundary then the ball would fall down. We test the learning simulations on 3 types of opponents: Static, Oﬀender, and Midﬁelder. Static-players do not move at all. Oﬀenders move randomly in x direction (i.e. they select randomly between Left and Right actions). Midﬁelders move randomly in all directions. States of the learner consists of absolute x and y position of ball and players, and status of ball from {Free, Picked } states, plus two state for left or right goal. This sums up to 3, 001, 252(= 2 × (7 × 5)4 + 2) states. Players act in parallel; so the next state of the learner when playing against oﬀenders or midﬁelders would be a stochastically selected state among 4 or 16 diﬀerent states, respectively. Each experiment consists of 100,000 episodes. Each episode starts by positioning the learner and ball in a random (and unoccupied) place. Each episode consists of maximum 1000 steps. Each step includes one movement by each player. An episode is ﬁnished if ball enters any goal or maximum number of steps passed. Upon delivering ball to the left goal the learner receives +1 reward. In case of wrong goal it receives -1 punishment. Otherwise it is punished with -0.01. All methods use SARSA with decaying -greedy policy. Learning rate is 0.1 and discount factor γ is 0.95. Exploration rate is initialized with 0.2 and decays by multiplying to decay factor 0.99999 after each episode. Leaf sample size is 360 (60 samples per action). For each state-abstraction method a range of conﬁdence thresholds are tested to ﬁnd the best settings; the one that minimizes the number of actions each learner executes from a random initial position to the goal. The number of actions is averaged over the last 10,000 episodes. This average shows how eﬀective has the agent learned to make goal in fewer moves. The results are averaged over 30 runs. Table 1 shows number of actions in each episode averaged over the last 10,000 episodes. UKS and UIGR are new versions of KS and IGR with unbi-

Reduction of Learning Time for Robots Using Automatic State Abstraction

89

Fig. 3. Number of actions (left) and leaves of the tree (right) during learning episodes in midﬁelder case, averaged per every 1000 episodes (non-accumulated) Table 1. Number of moves in each episode averaged over the last 10,000 episodes Player SARSA KS Static 10.85±.01 11.23±.2 Oﬀender 19.63 ± .06∗ 14.97±.57 Midﬁelder 668.36∗∗ 18.41±1.2 ∗ 14.72 after 1 million episodes

UKS IGR UIGR SMGR 11.35±.22 11.21±.05 11.23±.04 11.12±.03 14.91±.14 14.55±.15 14.58±.5 14.28±.07 17.39±.21 18.65±1.73 17.39±.24 16.84±.06 ∗∗ 18.92 after 15 million episodes

SANDS 11.1±.04 14.39±.09 16.95±.2

Table 2. Number of leaves in the learned trees SARSA Static 2452 Oﬀender 120052 Midﬁelder 3001252

KS 692±119 1476±404 1857±253

UKS 596±92 1331±78 1651±141

GR 703±35 1536±302 2263±504

UIGR 666±28 1242±23 1533±48

SMGR 516±10 1259±26 1436±78

SANDS 491±7 1282±22 1313±77

Table 3. Computation time of the algorithms in minutes Static Oﬀender Midﬁelder ∗ 16 min for 1

SARSA KS 1 14 11∗ 25 50∗∗ 35 million episodes

IGR SMGR SANDS 15 13 10 27 18 12 40 27 15 ∗∗ 370 min for 15 million episodes

ased sampling. Table 2 shows number of leaves in the learned trees(For SARSA it is the number of states). Conﬁdence intervals are computed with signiﬁcance level 0.05. Table 3 shows computation time in minutes 5 . Fig. 3 shows number of actions and leaves recorded during learning episodes in midﬁelder case.

5

Tests were done on a Dell R InspironTM 5150 laptop with [email protected] MHz.

90

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

5.1 Hierarchical Versus Flat RL From Table 1, it is clear that although SARSA shows better performance in static-player case but it has quite poor performance in other cases. Even we let the agent continue its learning episodes to 10 and 150 times more (1 and 15 million) for oﬀender and midﬁelder cases, respectively. The performance is still poor compared to hierarchical methods. By looking at time order of the HRL algorithms and ﬂat RL we observe that SARSA takes only one minute to learn the static-player case, with higher eﬃciency than all hierarchical algorithms. But when environment becomes more nondeterministic the required computation time increases exponentially and eﬃciency of the results is far from that of hierarchical approaches. An interesting result for HRL algorithms is that the required time increases in logarithmic order while the number of states increases exponentially. These results show that HRL have a great potential in decreasing learning time and number of trials in online robot learning tasks. 5.2 Unbiased Sampling Table 1 shows as environment becomes more and more nondeterministic, positive eﬀect of unbiased sampling becomes more clear. Although the average in UKS and UIGR is a little bit bigger than KS and IGR for static players, it is much smaller for midﬁelder case. Also number of leaves in the unbiased versions is always fewer than the regular versions; e.g. diﬀerence between IGR and UIGR in midﬁelder case is around 700 partitions. Conﬁdence intervals are always less for unbiased versions and this means that number of leaves in their generated trees does not vary a lot. 5.3 Eﬃciency of SANDS and SMGR Table 1 and Table 2 show in all cases SANDS and SMGR create trees with better performance than all other methods (even unbiased versions). The difference becomes more clear as the environment becomes more nondeterministic. In midﬁelder player case the diﬀerence between the new criteria and the old ones is about 2 actions per episode. The same is true for size of the generated trees. In almost all cases (except in UIGR, oﬀender player with negligible diﬀerence) SANDS and SMGR generate smaller trees. Figure 3 shows that SANDS and SMGR converge slower than the others. Although this can be problematic when state space is very big, in the sense that their performance does not increase very much for a long time, but it is expected that after a time, their performance anticipate the other criteria. Table 1 and Fig. 3 (Left) conﬁrm that splits in SANDS and SMGR are selected more carefully because not only their tree is the smallest, but also their performance is the best.

Reduction of Learning Time for Robots Using Automatic State Abstraction

91

SMGR shows a small enhancement in oﬀender and midﬁelder case compared to SANDS. It is due to a penalty term that SMGR adds to the splitting criterion which rejects some ineﬃcient splits. In static and oﬀender case it is not evident that which one generates fewer partitions. But in midﬁelder case SANDS generates trees with around 100 leaves less. 5.4 Computation Time As explained in Sect. 4.2, time order of SANDS and SMGR is less than KS and IGR. In static-player case the diﬀerence is not clear but in other cases SANDS decreases computation time to half compared to KS and even more compared to IGR. A part of the reduction in computation time is due to reduction in the order of sort algorithm in sample lists. Another part is due to creating smaller trees. This is important because it also aﬀects the speed of knowledge retrieval and the required memory of the trees. Another part is due to reduction in the number of actions in episodes.

6 Conclusion In this paper we formalized state abstraction as an optimization problem and by mathematical formulation we derived two new criteria for split selection. The new criteria adapt decision tree learning techniques to state abstraction. We showed in simulation that the new criteria outperform the existing criteria not only in eﬃciency and size of the learned trees but also in computation time. We believe that criteria, like KS and IGR, judge splits based only on their rank or probability of sample-values without taking their magnitude into account while; they are very informative for state abstraction. SANDS and SMGR mix probability and q-value of actions. We have tested other criteria like Information-Gain [19], Gini-Index [18] and Students T-test [10] but the results were with limited success, as reported in [10]. Students T-test has poor performance when variance over the splits gets close to zero because of a variance term in its divisor. Information-Gain and Gini-Index have poor performance in nondeterministic environments. Inability to revise the tree in non-stationary environments is a drawback of Univariate methods. Linear Multivariate [16] splits are perhaps more suitable candidates as they use a weighted sum of state-variables in split point. Adjusting these weights might be easier than reordering the nodes of the tree.

Acknowledgment This research is funded by the Future and Emerging Technologies programme (ISTFET) of the European Community, under grant IST-2001-35506. The information provided is the sole responsibility of the authors and does not reﬂect the Community’s opinion. The Community is not responsible for any use that might be made of data appearing in this publication. The Swiss participants to the project are supported under grant 01.0573 by the Swiss Government.

92

M. Asadpour, M.N. Ahmadabadi, and R. Siegwart

References 1. Sutton R.S. and Barto A.: Reinforcement Learning: An Introduction, MIT Press, Cambridge, MA (1996) 2. Uther W. and Veloso M.: The Lumberjack Algorithm for Learning Linked Decision Forests, In Proc. of PRICAI-2000 (2000) 3. Barto A.G. and Mahadevan S.: Recent Advances in Hierarchical Reinforcement Learning, Discrete Event Dynamic Systems, 13(4):41–77 (2003) 4. Sutton R.S., Precup D., and Singh S.: Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning, Artifcial Intelligence, 112:181–211 (1999) 5. Thrun S.B., and Schwartz A.: Finding structure in reinforcement learning, In Tesauro G., Touretzky D.S., and Leen T.(eds), Advances in Neural Information Processing Systems, pp.385–392, MIT Press, Cambridge, MA(1995) 6. Brooks R.A.: Achieving artiﬁcial intelligence through building robots, Technical Report A.I. Memo 899, AI Lab., MIT, Cambridge, MA (1986) 7. Dayan P. and Hinton G.E.: Feudal reinforcement learning, In Hanson S.J., Cowan J.D., and Giles C.L.(eds), Neural Information Processing Systems 5, San Mateo, CA, Morgan Kaufmann (1993) 8. Moore A.W.: The parti-game algorithm for variable resolution reinforcement learning in multidimensional state-spaces, In Cowan J.D., Tesauro G., and Alspector J.(eds), Advances in Neural Information Processing Systems, 6:711-718, Morgan Kaufmann (1994) 9. Chapman D., and Kaelbling L.P.: Input generalization in delayed reinforcement learning: An algorithm and performance comparisons, In Proc. of the 12th International Joint Conf. on Artiﬁcial Intelligence (IJCAI-91), pp.726-731 (1991) 10. McCallum A.: Reinforcement Learning with Selective Perception and Hidden State, PhD thesis, Computer Science Dept., Univ. of Rochester (1995) 11. Uther W.T.B. and Veloso M.M.: Tree based discretization for continuous state space reinforcement learning, In Proc. of the 15th National Conf. on Artiﬁcial Intelligence, AAAI-Press/MIT-Press, pp.769-774 (1998) 12. Uther W.T.B.: Tree Based Hierarchical Reinforcement Learning. PhD thesis, Dept. of Computer Science, Carnegie Mellon Univ., Pittsburgh, PA, USA (2002) 13. Jonsson A. and Barto A.G.: Automated state abstraction for options using the U-tree algorithm, In Advances in Neural Information Processing Systems: Proc. of the 2000 Conf., pp.1054–1060, MIT Press, Cambridge, MA (2001) 14. Dean T.,and Robert G.:Model minimization in markov decision processes, In Proc. AAAI-97, p.76 (1997) 15. Singh S.P., Jaakola T., and Jordan M.I.: Reinforcement learning with soft state aggregation, In Tesauro G., Touretzky D.S., and Leen T.K.(eds), Neural Information Processing Systems 7, MIT Press, Cambridge, MA (1995) 16. Yildiz O.T., Alpaydin E.: Omnivariate Decision Trees, IEEE Trans. on Neural Networks, 12(6):1539–1546 (2001) 17. Au M., Maire F.: Automatic State Construction using Decision Tree for Reinforcement Learning Agents, International Conf. on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA), Gold Coast, Australia (2004) 18. Breiman L., Friedman J., Olshen R., Stone C.: Classiﬁcation and Regression Trees, Wadsworth, Paciﬁc Grove, CA (1984) 19. Quinlan R.: Induction of decision trees, Machine Learning, 1:81–106 (1986)

Eﬃcient Failure Detection for Mobile Robots Using Mixed-Abstraction Particle Filters Christian Plagemann, Cyrill Stachniss, and Wolfram Burgard University of Freiburg Georges-Koehler-Allee 79110 Freiburg, Germany {plagem,stachnis,burgard}@informatik.uni-freiburg.de Summary. In this paper, we consider the problem of online failure detection and isolation for mobile robots. The goal is to enable a mobile robot to determine whether the system is running free of faults or to identify the cause for faulty behavior. In general, failures cannot be detected by solely monitoring the process model for the error free mode because if certain model assumptions are violated the observation likelihood might not indicate a defect. Existing approaches therefore use comparably complex system models to cover all possible system behaviors. In this paper, we propose the mixed-abstraction particle ﬁlter as an eﬃcient way of dealing with potential failures of mobile robots. It uses a hierarchy of process models to actively validate the model assumptions and distribute the computational resources between the models adaptively. We present an implementation of our algorithm and discuss results obtained from simulated and real-robot experiments.

1 Introduction Whenever mobile robots act in the real world, they are aﬀected by faults and abnormal conditions. Detecting such situations and allowing the robot to react appropriately is a major precondition for truly autonomous vehicles. While the applied techniques need to be able to reliably detect rare faults, the overall estimation process under error-free conditions should not be substantially more complex compared to systems that are optimized for the normal operational mode. Separate monitoring processes that use more complex models to cover all possible system behaviors introduce an unnecessary high computational load. In this paper, we introduce mixed-abstraction particle ﬁlters as an eﬀective means for adaptively distributing the computational resources between diﬀerent system models based on the estimated validity of their speciﬁc model assumptions. The term ”fault detection” is commonly referred to as the detection of an abnormal condition that may prevent a functional unit from performing a

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 93–107, 2006. © Springer-Verlag Berlin Heidelberg 2006

94

C. Plagemann, C. Stachniss, and W. Burgard

Fig. 1. The left image depicts a simulated robot before colliding with an obstacle which is not detected by its sensors. The right photograph shows a real mobile robot that collides with an undetected glass door while moving on its planned trajectory to the neighboring room.

required function [11]. Most works in the fault detection and isolation literature deal with internal faults such as defects in hardware or software. For the mobile robot domain, we apply the same nomenclature to external inﬂuences like collisions or wheel slip since their eﬀects are similar to those of internal defects and the resulting models have the same structure. As an illustrating scenario, consider a mobile robot equipped with wheel encoders, a laser range ﬁnder, and a suﬃciently accurate map of the environment. In the fault-free case, the position of the robot can be tracked using a standard tracking algorithm such as a Kalman ﬁlter or a particle ﬁlter with a simplistic odometry-based motion model, which is formally given in Section 3.3. In odometry-based models, the next system state xt is directly predicted from the odometry, which is the measurement ot obtained from the wheel encoders. Although such models allow us to evaluate diﬀerent state hypotheses by weighting them using exteroceptive measurements, e.g., using a laser range measurement lt , they do not directly allow us to detect collisions with obstacles that cannot be perceived by the sensors of the robot. This is due to the fact that when the robot stops moving, its wheel encoders do not record any motion, which is perfectly consistent with the recorded laser measurements. Therefore, no ﬁlter degradation occurs and there is no possibility to detect such faults inside the ﬁlter. One typical solution to overcome such problems is to compare the estimated trajectory with the planned one on a higher system level. As major drawbacks of such an approach, one cannot infer the actual cause for the deviation from the planned trajectory and the system architecture is complicated by the stronger connection between the planning and tracking module. An alternative solution is to consider the actual motion commands that have been sent to the motors instead of just using the wheel encoder readings. However, this makes the system model substantially more complex and

Eﬃcient Failure Detection for Mobile Robots

95

the predictions, which are now based on motor currents and accelerations, less accurate. In our experiments, we observed that such a model is around 32 times slower to compute than the odometry-based model. It is important to note that the odometry-based model makes the implicit assumption, that the wheel encoder measurements reﬂect the intended motion. If this assumption is violated, the standard estimation technique cannot be used to estimate the joint probability p(x, f ) anymore, where x stands for the pose of the robot and f indicates the failure state. In this paper, we propose to make the model assumptions explicit and to build a model abstraction hierarchy. We present the mixed-abstraction particle ﬁlter algorithm that uses such a hierarchy to direct computational resources to the most eﬃcient model whose assumptions are met. In this way, it minimizes the computational load while maximizing the robustness of the system.

ot xt-1

Ut-1 xt

xt-1 lt

xt ot

lt

Fig. 2. Arbitrary model hierarchy (left) with an unrestricted model M0 , several restricted models M1 , M2 , M3 as well as the speciﬁc assumptions Ai→j that restrict the state spaces respectively. Two models for the same physical process: the standard odometry-based model (e.g. M1 ) that uses the odometry measurements ot as controls (middle) and the laser measurements lt as sensor inputs. A less restricted model (e.g. M0 , on the right) that includes the actual motion commands ut as controls.

The paper is organized as follows. After the discussion of related work, we present our mixed-abstraction particle ﬁlter algorithm and its application to monitoring mobile robots in Section 3. Finally, Section 4 presents experimental results.

2 Related Work Model-based diagnosis has been approached from within the AI community using symbolic reasoning with a focus on large systems with many interacting components and from the control theory community concentrating on fewer components with complex dynamics and higher noise levels [8]. With symbolic approaches, the system is typically assumed to move between discrete

96

C. Plagemann, C. Stachniss, and W. Burgard

steady states [17]. Here, diagnosis is often based on system snapshots without a history. Krysander [8] proposed a hybrid model consisting of discrete fault modes that switch between diﬀerential equations to describe the system behavior. The diagnosis system is designed for large systems with low noise levels, where instantaneous statistical tests are suﬃcient to identify a faulty component. As Dearden and Clancy [3] pointed out, the close coupling between a mobile system with its environment makes it hard to apply discrete diagnosis models directly, because extremely complex monitoring components would have to be used. A more robust and eﬃcient way is to reason directly on the continuous sensor readings. As a result, probabilistic state tracking techniques have been applied to this problem. Adopted paradigms range from Kalman ﬁlters [16] to particle ﬁlters in various modelings [1, 3, 12]. Particle ﬁlters represent the belief about the state of the system by a set of state samples, which are moved when actions are performed and re-weighted when sensor measurements are integrated (see Dellaert et al. [4]). In particle ﬁlter based approaches to failure diagnosis, the system is typically modeled by a dynamic mixture of linear processes [2] or a non-linear Markov jump process [5]. Benanzera et al. [1] combine consistency-based approaches, i.e. the Livingstone system, with particle ﬁlter based state estimation techniques. Verma et al. [15] introduce the variable resolution particle ﬁlter for failure detection. Their approach is similar to ours in that they build an abstraction hierarchy of system models. The models of consideration build a partition of the complete state space and the hierarchy is deﬁned in terms of behavioral similarity. Our focus in contrast lies on switching between overlapping system models for certain parts on the state space. Our model hierarchy is therefore based on eﬃciency diﬀerences and explicit model assumptions about the system state. The two approaches should therefore be seen as complementary rather than alternatives. Other approaches that deal with the time eﬃciency of particle ﬁlters include Kwok et al. [10] in which real-time constraints are considered for single system models or techniques in which a Rao-Blackwellized particle ﬁlter is used to coordinate multiple models for tracking moving objects [9].

3 Particle Filters for Sequential State Estimation A mobile robot can be modeled as a dynamic system under the inﬂuence of issued control commands ut and received sensor measurements zt . The temporal evolution of the system state xt can be described recursively using the formalism of the so called Bayes ﬁlter

Eﬃcient Failure Detection for Mobile Robots

97

Algorithm 1 Particle ﬁlter(Xt−1 , ut−1 , zt ) 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11:

X t = Xt = ∅ for m = 1 to M do [m] [m] sample xt ∼ p(xt | ut−1 , xt−1 ) [m] [m] wt = p(zt | xt ) [m] [m] X t = X t + x t , wt end for for m = 1 to M do [i] draw particle i with probability ∝ wt [i] add xt to Xt end for return Xt

p(xt | z0:t , u0:t−1 ) = = ηt

p(xt | zt , ut−1 , xt−1 ) p(xt−1 | z0:t−1 , u0:t−2 ) dxt−1 p(zt | xt ) observation model

(1)

p(xt | ut−1 , xt−1 ) p(xt−1 | z0:t−1 , u1:t−2 ) dxt−1 . (2) motion model

recursive term

The term ηt is a normalizing constant ensuring that the left-hand side sums up to one over all xt . With Equation 1, we assume Markovian dependencies, namely that xt only depends on the most recent measurement zt and control command ut−1 given knowledge about the preceding state of the system xt−1 . Particle ﬁlters are an implementation of the Bayes ﬁlter and can be used to eﬃciently estimate the posterior p(xt ) in a sequential manner. Here, the pos[m] [m] terior is represented by a discrete set of weighted samples X = { xt , wt }. With the sampled representation, the integral in Equation 2 simpliﬁes to a ﬁnite sum over the samples resulting from the previous iteration. The motion model and the observation model can be applied directly to move and weight the individual samples respectively. Algorithm 1 formulates the standard particle ﬁltering approach [14]. In Algorithm 1, the state of the system at time t is represented by a set Xt [m] of state samples xt . In Line 3, we perform a state prediction step using [m] the external motion command ut−1 and the motion model p(xt | ut−1 , xt−1 ). Line 4 incorporates the current sensor measurement zt by re-weighting the [m] state samples according to the measurement model p(zt | xt ). From Line 7 to 10, a resampling step is performed to concentrate the samples on highprobability regions of the state space. We refer to Thrun et al. [14] for details about the resampling step and its eﬃcient implementation. For the odometry-based model that treats the odometry measurements as controls, we have ut−1 = ot and zt = lt , where ot is the odometry measurement and lt is a perception of the environment. For the dynamic model depicted

98

C. Plagemann, C. Stachniss, and W. Burgard

in the right diagram of Figure 2, ut−1 is the actual control command and zt = ot , lt . Both models are described in more detail in Section 3.3. 3.1 Process Model Hierarchy A fundamental problem in science and engineering is choosing the right level of abstraction for a modeled system. While complex and high-dimensional models promise high estimation accuracy, models of less complexity are often signiﬁcantly more eﬃcient and easier to construct. In the area of mobile robotics, the accuracy-eﬃciency trade-oﬀ is an important issue, since on the one hand, computational resources are strictly limited in online problems and on the other hand, estimation errors have to be avoided to prevent serious malfunctioning. We therefore propose an online model selection algorithm with adaptive resource allocation based on the Bayesian framework. An abstraction hierarchy for process models is given by the speciﬁc assumptions that the diﬀerent models make about the world (compare Figure 2). We deﬁne the model abstraction hierarchy as an acyclic directed graph with the diﬀerent system models Mi as nodes and their model assumptions Ai→j as edges, leading from the more general model i to a more restricted one j. A model assumption Ai→j is deﬁned as a binary function on the state space of the unrestricted model Mi . As an example, consider the process model for a mobile robot that should be able to continuously localize itself in a given map. A general model M0 would include the the pose of the robot x and additionally take physical factors like ground friction, tire pressure, load balance, motor characteristics, etc. into account and treat those as additional state variables in a state vector f . In most situations, however, it is quite common and reasonable to assume a simpler model M1 , where these additional variables are constant and do not need to be estimated during operation. Formally, the state space of M1 is therefore {x, f | f = const}, which is a projection of the more general space {x, f } of model M0 . The assumption A0→1 would in this case be deﬁned as true : f = const A0→1 (x, f ) := false : f = const. It is important to note that the validity of an assumption can only be tested in a less restricted state space, where this assumption is not made. In practice, this means that we have to test for every edge in the model abstraction graph the associated assumption using the more general model. As a measure for the validity of an assumption A˜ at time t, we use the ratio ˜ := vt (A)

A

p(zt | x[m] )

p(zt | x[m] )

1 |A| 1 |X |

,

(3)

˜ [m] )} of all particles X for which where A is deﬁned as the subset {x[m] | A(x A˜ is valid. More informally, we compute the amount of evidence in favor of

Eﬃcient Failure Detection for Mobile Robots

99

Algorithm 2 Mixed-Abstraction Particle ﬁlter 1: Calculate samples for the unrestricted model Mi until the assumption Ai→j (˜ xt ) is valid for a minimal number of samples x ˜t 2: Build a ﬁrst estimate of vt (Ai→j ) according to Equation 3 3: repeat 4: if vt (Ai→j ) >= Θ then 5: Calculate samples for Mj 6: else 7: Calculate samples for Mi 8: end if 9: until either Mi received enough samples or (vt (Ai→j ) >= Θ and Mj received enough samples)

a restricted state space relative to the unrestricted case. The quantity vt (A˜) is based on the current approximation of the posterior distribution by the particle ﬁlter. 3.2 Adaptive Model Selection To adaptively switch between alternative system models, the validity of the model assumptions have to be estimated online and computational resources have to be distributed among the appropriate models. The distribution of resources is done by increasing or decreasing the number of particles for the diﬀerent models. To achieve this, we apply the following algorithm that takes as input the model abstraction hierarchy graph deﬁned in the previous section. When a new measurement zt is obtained, the mixed abstraction particle ﬁlter algorithm draws samples from the particle set representing the current posterior Xti−1 for all system models i, incorporates the measurement, and builds new posterior distributions Xti . The key question in this update step is which model posterior Xti should receive how many samples. We base this decision on the estimated validity of the model assumptions Ai→j . If the estimated quantity vt (Ai→j ) drops below a predeﬁned threshold Θ, we sample into the more complex model Mi and otherwise prefer the more eﬃcient one Mj . This decision is made repeatedly on a per particle basis until a model has received enough samples and all its assumptions are validated. In each iteration, we start with the most unrestricted model Mi and perform the following steps for each of its outgoing edges Ai→j . In the update steps mentioned above, the samples are taken from the j , if assumption Ai→j was valid in the previous posterior distributions Xt−1 i previous step and from Xt−1 otherwise. When all outgoing edges Ai→j of model Mi have been processed in the described manner, the same update is applied to models Mj further down in the hierarchy until either the leaf nodes have been processed or one of the assumptions did not receive enough evidence to justify further model simpliﬁcations. Several quantities like the numbers of samples necessary for each model (Line 9) or the validity threshold Θ (Line 4) have to be determined in an oﬄine

100

C. Plagemann, C. Stachniss, and W. Burgard

learning step. For the experimental results reported below, we optimized these values on a set of representative trajectories, recorded from real and simulated robots. To recapitulate, the mixed abstraction particle ﬁlter estimates the system state by running several particle ﬁlters in parallel, each using a diﬀerent system model. Samples are assigned applying the following rule. For each two alternative system models, the simpler one is prefered as long as there is positive evidence for the validity of the corresponding model assumption. 3.3 Motion Models for Mobile Robots The standard odometry-based motion model for a wheeled robot estimates the posterior p(xt | xt−1 , ot ) about the current pose xt based on the previous pose xt−1 and the wheel encoder measurement ot . A popular approach [6] to represent the relative movement is to use three parameters, an initial rotation α, a translation d, and a second rotation β as illustrated in Figure 3. Typically, one uses a Gaussian distribution for each of these parameters to model the noise. measured pose

β’

β

d’

initial pose α

α’

d

final pose path

Fig. 3. Parameters of the standard odometry-based motion model.

Under the inﬂuence of events like the collision with an obstacle or wheel slippage, the odometry-based model is not applicable anymore since the wheel encoder measurements do not provide useful information about the actual motion. To handle such situations, we construct an alternative model, termed dynamic motion model, that depends on the actual motion commands that were sent to the motors of the robot. This model includes the geometry of the robot and its physical attributes like mass and moment of inertia. For each wheel, we compute the inﬂuence of the velocity command on the translational and rotational energy of the robot. In this representation, we can directly incorporate the eﬀects of collisions, slippage, and deﬂating tires. We then convert the energy state of the system to its speed and obtain a state prediction for the next time step. It is important to note that the dynamic motion model is not designed for the failure states only. Rather it is able to deal with normal conditions

Eﬃcient Failure Detection for Mobile Robots

101

as well. It is therefore considered as more general as the standard odometrybased model in our model abstraction hierarchy. The assumption placed on the system state by the odometry-based model is, that there are no external inﬂuences like collisions, slippage, etc.

4 Experiments 4.1 Quantitative Evaluation Using a Simulator To quantitatively evaluate our system, we performed several simulation experiments and compared the results to the known ground truth. We used the high-ﬁdelity simulator Gazebo [7], in which physics and motion dynamics are simulated accurately using the Open Dynamics Engine [13]. In several practical experiments carried out with real robots, we experienced the Gazebo simulator as well suited for studying the motion dynamics of robots even under highly dynamic events like collisions. For example, we did not have to change any system parameters when we ported our system from the simulator to the real robot. To demonstrate how the proposed algorithm coordinates multiple particle ﬁlters that have been designed independently, we confronted the system with two diﬀerent faults within one scenario. A simulated ActivMedia Pioneer 2DX robot (see the left image of Figure 1) was placed in the corridor of a 3D oﬃce environment. We manually steered the robot through this environment. On its path, it encountered a collision with an undetectable object. After that, its left tire started to deﬂate. Four ﬁlters were used independently to track the state of the system. The ﬁrst ﬁlter was based on the standard odometry-based motion model described in Section 3.3. The second ﬁlter used the dynamic motion model described in the previous section and also included a model for collision faults. The third model was also based on the dynamic motion model but was capable of dealing with deﬂating tires. Finally, the forth ﬁlter was the proposed mixed-abstraction ﬁlter that combined all of the ﬁlters described above. Figure 4 depicts the true trajectory as well as the tracking results obtained with the individual ﬁlters overlayed on a map built from laser measurements in a real building. The three arrows in the diagram mark the following three events: a collision with an undetected glass door (1), a point where the robot stopped backing oﬀ and turned around (2), and a point where the left tire of the robot started losing air (3). As can be seen from the ﬁgure, the ﬁlter that is able to deal with deﬂating tires diverges immediately after the collision at point 1. Since the ﬁlter able to deal with collisions cannot deal with deﬂating tires, it diverges at point 3. The odometry-based model keeps track of the robot despite the collision at point 1, however it is not aware of any fault at this point. The combined ﬁlter in contrast succeeds in tracking the robot despite of both incidents.

102

C. Plagemann, C. Stachniss, and W. Burgard

The middle and lower image of Figure 4 plot the internal states of the specialized detectors within the mixted-abstraction ﬁlter. These values reﬂect the belief of the system about the presence of certain faults. The middle image plots the relative number of particles in the fault mode of the collision detector over time. As can be seen, this number raises signiﬁcantly shortly after the collision as well as after the full stop at point 2. After the robot had been intenionally stopped there, the system cannot know whether an obstacle is in its way or not. The lower image of Figure 4 shows the evolution of the relative number of particles in the fault mode of the deﬂation detector. Since the collision at point 1 has been handled by the collision detection within the mixed-abstraction ﬁlter, this ﬁlter does not switch to a failure mode until point 3. At that point, the ﬁlter switches into its failure mode and in this way enables the mixed-abstraction ﬁlter to keep track of the pose of the robot. 4.2 Analyzing the Gain in Eﬃciency In this experiment, we quantatively evaluate the gain in eﬃciency that we achieve by dynamically distributing samples between the individual ﬁlters. In the modeled scenario, a simulated robot follows a trajectory and collides twice with an obstacle, which is too small to be detected by its laser range ﬁnder. After the collisions, the robot backs oﬀ and continues its task. The top diagram of Figure 5 shows the trajectory of the vehicle and the locations where the algorithm correctly detected a collision. The other two diagrams illustrate the failure detection process of the same simulation run. The bar at the bottom indicates the true time stamps of the faults. Whereas the plot in the second row depicts the relative likelihood for a collision as deﬁned in Equation 3, the curve plotted in the third row gives the times needed for the individual iterations of the particle ﬁlter. Table 1 gives the results of a comparison of our adaptive model-switching approach to three other implementations, where only single models were used for state estimation and fault detection. The results are averaged over the full trajectories of 100 runs per implementation. The implementations considered here are realized on the basis of two models. Whereas model M0 is the complex model that considers the actual motion commands and therefore is able to track the position as well as the failure state, model M1 is the standard odometry-based system model, which is able to track the position of the robot reliably with low time requirements, but cannot detect the collisions. The ﬁrst implementation is based on model M1 with 20 particles, the second is model M1 with 200 particles. While the third implementation is based on model M0 with 300 particles, the forth one is the mixed-abstraction particle ﬁlter combining implementation one and three. The common task of all implementations was to track the position of the robot along a trajectory on which the robot encountered two undetected collisions after 8.2 and 28.9 seconds. A typical estimate of the trajectory generated by the mixed-abstraction ﬁlter including markings for detected collisions is depicted in the left diagram of

Eﬃcient Failure Detection for Mobile Robots

1 0.5 0

1 0.5 0

103

Collision Detector 1 2 3 22 24 26 28 30 32 34 36 38 40 Simulation time [sec] Deflation Detector 1 2 3 22 24 26 28 30 32 34 36 38 40 Simulation time [sec]

Fig. 4. Results of an experiment with two fault-events (top). The collision of the robot with a glass door is marked with the arrow “1”, the point where it stopped backing of and turned is marked with “2”, and at point “3” the left tire of the robot started losing air. The diagram in the middle plots the relative number of the particles in the fault mode of the collision detector. The lower diagram shows the corresponding number for the deﬂation detector.

Figure 5. The lower right image of Figure 5 plots the value of vt (A) over time. As can be seen from the ﬁgure, the evidence for a fault substantially increases at the time of the incidences. The upper right image in the same ﬁgure plots the CPU time used by the mixed-abstraction ﬁlter. It nicely shows that the computation time is only high when the evidence of a failure has increased. Table 1 shows average values obtained from the 100 test runs for each implementation. Whereas model M1 was never able to detect a failure, M0

104

C. Plagemann, C. Stachniss, and W. Burgard 17

Collision detected Estimated trajectory

Estimated y-position

16.5

t=8.2 sec

16 15.5 15 14.5

t=50.0 sec t=0.0 sec

14

Rel. likelihood

13.5

7

8

8 7 6 5 4 3 2 1 0

9 10 11 12 Estimated x-position

13

14

Relative likelihood: collision

0 Time per iteration [ms]

t=28.9 sec

10

20

30

40

50

Simulation time [sec] 35 30 25 20 15 10 5 0

Time per iteration

0

10

20

30

40

50

Simulation time [sec]

Fig. 5. A trajectory followed by a simulated robot (ﬁrst row) with marks at positions where the evidence for a collision was high. The plot in the second row depicts the relative likelihood for collisions. The plot also shows the ground truth (the bar at the bottom). The last plot shows the time needed for each iteration (third row).

as well as our adaptive switching algorithm detected all failures equally well. However, our adaptive model required substantially less computation time compared to M0 alone using 300 particles. 4.3 Evaluation on a Real Robot We also tested our system on a real ActivMedia Pioneer 2DX robot in an ofﬁce environment. The right image of Figure 1 depicts the experimental setup. Three positions of the robot were manually cut from a recorded video and overlayed on one image to illustrate the process. The robot planned a path to the neighboring room on the right-hand side of the corridor. While executing the planned trajectory, the robot could not detect the glass door that

Eﬃcient Failure Detection for Mobile Robots

105

Table 1. Results of a series of simulation runs using diﬀerent system models for state estimation. The results are averaged over the complete trajectories of 100 runs per model. System model

Failure Detection Average time Average estimation error rate per iteration Position Orientation M1 : standard odometry 0% 0.67 ms 0.13 m 3.6◦ 20 particles M1 : standard odometry 0% 5.83 ms 0.13 m 3.4◦ 200 particles M0 : dynamic 100 % 10.10 ms 0.11 m 5.8◦ 300 particles adaptive-switching: 100 % 3.42 ms 0.12 m 3.9◦ M1 : 20 particles M0 : 300 particles

blocked its path and thus collided with the wooden part of the door. In this situation, the standard odometry-based system model used for localization does not indicate a defect, because the wheel encoders report no motion, which is perfectly consistent with the laser measurements. The left diagram of Figure 6 gives the evolution of the observation likelihoods for the standard model, which stays nearly constant. In contrast to this, the proposed mixed abstraction particle ﬁlter detects that the model assumption of the standard model is not valid anymore after the collision with the door and switches to the more complex system model. The right diagram of Figure 6 visualizes this process. For the sake of clarity, we plotted the estimated validity of the negated model assumption, which can be interpreted as the evidence against the assumption that no collision has occurred. The upper curve corresponds to the time needed per iteration. Note that the required computational resources only slightly exceed those of the standard odometry-based model (see left diagram for a comparison). Only in the failure case, the runtime increases seriously since the more complex model requires substantially more particles. Please also note, that the runtime goes back to the original value after the robot has backed oﬀ and the model assumption of the simpliﬁed model is valid again.

5 Conclusion This paper presents an eﬃcient approach to estimate the state of a dynamic system including its failures. Complex models with high computational requirements are needed in order to detect and track unusual behaviors. We therefore proposed a mixed-abstraction particle ﬁlter which distributes the computational resources in a way that failure states can be detected and tracked but at the same time allows us an eﬃcient estimation process in case

C. Plagemann, C. Stachniss, and W. Burgard 1

Time per iteration [ms]

0.5 0

0

5

10

15

20

25

30

Time per iteration [ms]

Time per iteration [ms]

106

40

Time per iteration [ms]

30 20 10 0

0

5

10

0.03

Observation likelihood

0.02 0.01 0

0

5

10

15 Timestamp [sec]

15

20

25

30

Timestamp [sec] Rel. likelihood

Observation likelihood

Timestamp [sec]

20

25

30

10 8 6 4 2 0

Relative likelihood (collision)

0

5

10

15

20

25

30

Timestamp [sec]

Fig. 6. In an experiment with a real robot, the observation likelihood of the standard system model (lower left) does not indicate the collision with a glass door at time t = 22 seconds. The mixed-abstraction particle ﬁlter (right) detects the fault without needing substantially more computational resources in the fault-free case (upper diagrams).

the systems runs free of faults. To achieve this, we apply a process model hierarchy which allows us to model assumptions that hold for the fault-free case but not in general. In several experiments carried out in simulation and with real robots, we demonstrated that our technique is well-suited to track dynamic systems aﬀected by errors. Our approach allows us to accurately track diﬀerent failure states and at the same time is only marginally slower in case the system is running free of faults. We believe that our approach is not limited to the failure detection problem and can also be advantageous for various state estimation tasks in which diﬀerent system models have to be used to correctly predict the behavior of the system under varying conditions.

Acknowledgments This work has partly been supported by the EC under contract number FP6004250-CoSy, by the German Science Foundation (DFG) under contract number SFB/TR-8 (A3) and the German Federal Ministry of Education and Research (BMBF) under contract number 01IMEO1F.

References 1. E. Benazera, R. Dearden, and S. Narasimhan. Combining particle ﬁlters and consistency-based. In 15th International Workshop on Principles of Diagnosis, Carcassonne, France, 2004. 2. N. de Freitas, R. Dearden, F. Hutter, R. Morales-Menendez, J. Mutch, and D. Poole. Diagnosis by a waiter and a mars explorer. In Invited paper for Proceedings of the IEEE, special issue on sequential state estimation, 2003.

Eﬃcient Failure Detection for Mobile Robots

107

3. R. Dearden and D. Clancy. Particle ﬁlters for real-time fault detection in planetary rovers. In Proceedings of the Thirteenth International Workshop on Principles of Diagnosis, pages 1–6, 2002. 4. F. Dellaert, D. Fox, W. Burgard, and S. Thrun. Monte carlo localization for mobile robots. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), Leuven, Belgium, 1998. 5. J.N. Driessen and Y. Boers. An eﬃcient particle ﬁlter for nonlinear jump markov systems. In IEEE Sem. Target Tracking: Algorithms and Applications, Sussex, UK, 2004. 6. J.-S. Gutmann, W. Burgard, D. Fox, and K. Konolige. An experimental comparison of localization methods. In Proc. of the IEEE/RSJ InternationalConference on Intelligent Robots and Systems, 1998. 7. N. Koenig and A. Howard. Design and use paradigms for gazebo, an opensource multi-robot simulator. technical report. Technical report, USC Center for Robotics and Embedded Systems, CRES-04-002, 2004. 8. M. Krysander. Design and Analysis of Diagnostic Systems Utilizing Structural Methods. PhD thesis, Link¨ oping University, Sweden, 2003. 9. C. Kwok and D. Fox. Map-based multiple model tracking of a moving object. In RoboCup 2004: Robot Soccer World Cup VIII, pages 18–33, 2004. 10. C. Kwok, D. Fox, and M. Meila. Real-time particle ﬁlters. In Suzanna Becker, Sebastian Thrun, and Klaus Obermayer, editors, Advances in Neural Information Processing Systems 15 (NIPS), pages 1057–1064. MIT Press, 2002. 11. N. Leveson. Safeware : System Safety and Computers. Addison-Wesley Pub Co., Reading, Mass., 1995. 12. B. Ng, A. Pfeﬀer, and R. Dearden. Continuous time particle ﬁltering. In Proceedings of the 19th IJCAI, Edinburgh, 2005. 13. R. Smith. Open dynamics engine. http://www.q12.org/ode/ode.html, 2002. 14. S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics. MIT Press, 2005. 15. V. Verma, S. Thrun, and R. Simmons. Variable resolution particle ﬁlter. In Georg Gottlob and Toby Walsh, editors, IJCAI-03, Proceedings of the Eighteenth International Joint Conference on Artiﬁcial Intelligence, Acapulco, Mexico, August 9-15, 2003, pages 976–984. Morgan Kaufmann, 2003. 16. R. Washington. On-board real-time state and fault identiﬁcation for rovers. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 1175–1181, 2000. 17. B.C. Williams and P.P. Nayak. A model-based approach to reactive selfconﬁguring systems. In Jack Minker, editor, Workshop on Logic-Based Artiﬁcial Intelligence, Washington, DC, College Park, Maryland, 1999. University of Maryland.

Optimization and Fail–Safety Analysis of Antagonistic Actuation for pHRI Gianluca Boccadamo, Riccardo Schiavi, Soumen Sen, Giovanni Tonietti, and Antonio Bicchi Interdepartmental Research Centre “E. Piaggio” via Diotisalvi 2, Faculty of Engineering University of Pisa, Italy @ing.unipi.it Summary. In this paper we consider some questions in the design of actuators for physical Human-Robot Interaction (pHRI) under strict safety requirements in all circumstances, including unexpected impacts and HW/SW failures. We present the design and optimization of agonistic-antagonistic actuation systems realizing the concept of variable impedance actuation (VIA). With respect to previous results in the literature, in this paper we consider a realistic physical model of antagonistic systems, and include the analysis of the eﬀects of cross–coupling between actuators. We show that antagonistic systems compare well with other possible approaches in terms of the achievable performance while guaranteeing limited risks of impacts. Antagonistic actuation systems however are more complex in both hardware and software than other schemes. Issues are therefore raised, as to fault tolerance and fail safety of diﬀerent actuation schemes. In this paper, we analyze these issues and show that the antagonistic implementation of the VIA concept fares very well under these regards also.

1 Introduction One of the goals of contemporary robotics research is to realize systems which operate with delicacy in environments they share with humans, ensuring their safety despite any adverse circumstance [4]. These may include unexpected impacts, faults of the mechanical structure, sensors, or actuators, crashes or malfunctional behaviours of the control software [6, 7, 1, 5]. A recent trend in robotics is to design intrinsically safe robot arms by introducing compliance at their joints. The basic idea of this approach is that compliant elements interposed between motors and moving links help prevent the (heavy) reﬂected inertia of actuators from concurring to damage in case of impacts. Introducing compliance, on the other hand, tends to reduce performance of the arm. Some approaches in the direction of minimizing the perfor-

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 109–118, 2006. © Springer-Verlag Berlin Heidelberg 2006

110

G. Boccadamo et al.

mance loss while guaranteeing safety in case of impacts have been presented in the recent literature (see e.g. [2]). Among these, a method was proposed in [2] consisting in varying the compliance of the joint transmission mechanism while moving the arm. This so-called Variable Stiﬀness Transmission (VST) technique, and its generalization in the Variable Impedance Actuation (VIA) concept, have been shown to be capable in theory of delivering better performance than purely passive compliance and other techniques. In its formulation, however, the VIA concept in [2] used a rather abstract model of actuator and transmission, whereby the impedance could be directly controlled to desired values in negligible time. In this paper, we consider a more realistic model of an actuation system implementing the idea, which is based on the use of two actuators and nonlinear elastic elements in antagonistic arrangement. The antagonistic solution has several advantages, and has been used in many robotic devices before (see e.g. [3, 11, 14, 13, 8]), in some cases because of biomorphic inspiration. However, to the best of our knowledge the introduction of nonlinear springs to achieve variable stiﬀness in real time (i.e., during diﬀerent phases of each motion act) was not a motivation for earlier work with the purpose of guaranteeing safety. In this paper, we consider the implementation of the VIA concept by means of antagonistic actuation, discuss the role of cross-coupling between antagonist actuators, and apply optimization methods to choose parameters which are crucial in its design. We show that antagonistic systems can implement eﬀectively the VIA concept, and their performance compares well with other possible approaches. Antagonistic actuation systems however are more complex in both hardware and software than other schemes. Issues are therefore raised as to whether safety is guaranteed under diﬀerent possible failure modes. In the paper, we also analyze these issues and show that the antagonistic implementation of the VIA concept fares very well under these regards also.

2 Antagonistic Actuation as a VIA System In [2] it was shown that an ideal VIA mechanism (depicted in ﬁg. 1-a) can eﬀectively recover performance of mechanisms designed to guarantee safety of humans in case of impact. The basic idea is that a VIA mechanism can be controlled according to a stiﬀ-and-slow/fast-and-soft paradigm: namely, to be rather stiﬀ in the initial and ﬁnal phases of motion, when accuracy is needed and velocity is low, while choosing higher compliance in the intermediate, high-velocity phase, where accuracy is typically not important. Low stiﬀness implies that the inertia of the rotors does not immediately reﬂect on the link in case of impacts, thus allowing smoother and less damaging impacts. Such arguments were supported in [2] by a detailed mechanism/control co-design optimization analysis, based on the solution of the so-called safe brachistochrone

Optimization and Fail–Safety of Antagonistic Actuation for pHRI

a)

111

b)

Fig. 1. The concept of Variable Impedance Actuation (a) and a possible implementation by means of antagonistic actuators (b). Eﬀective rotor inertias are coupled to the link inertia through nonlinear springs.

a)

b)

Fig. 2. An experimental implementation of an antagonistic VIA actuator (a) and its conceptual scheme (b).

problem, i.e. a minimum time control problem with constraints on the maximum acceptable safety risk at impacts. The model considered ﬁg. 1-a, however, uses direct variations of impedance, which is not physically realizable. A possible implementation of the concept via an antagonistic mechanism is depicted in ﬁg. 1-b. Practical implementations of antagonistic VIA systems may assume more general conﬁgurations than the one in ﬁg. 1-b. For instance, in the prototype of an antagonistic VIA system depicted in ﬁg. 2-a ([10]), the two actuators act through a nonlinear elastic element on the link, but they are also connected to a third elastic element cross-coupling the actuators. Questions we consider in this section are the following: is the stiﬀ-andslow/fast-and-soft control paradigm still valid, and are the good safety and performance properties of the ideal VIA device ﬁg. 1-a retained by an antagonistic implementation as in ﬁg. 1-b? What is the role of cross-coupling elastic elements as in ﬁg. 2-b in antagonistic VIA actuators? To answer these questions, we use again the analysis of solutions to the safe brachistochrone problem, which consists in ﬁnding the optimal motor torques τ1 , τ2 which drive the link position xmov between two given conﬁgurations in minimal time, subject to the mechanism’s dynamics, motor torque limits, and safety constraints. This problem is formalized for the antagonistic mechanism of ﬁg. 1-b as

112

G. Boccadamo et al.



T

minτ 0 1 dt Mrot1 x ¨rot1 + φ1 (xrot1 , xmov ) = τ1  Mrot2 x ¨rot2 − φ2 (xrot2 , xmov ) = τ2 Mmov x ¨link − φ2 (xrot2 , xmov ) − φ1 (xrot1 , xmov ) = 0 |τ1 | ≤ U1,max |τ | ≤ U2,max  2 HIC(x˙ mov , x˙ rot1 , x˙ rot2 , φ1 , φ2 ) ≤ HICmax ,

(1)

where Mmov , Mrot1 , Mrot2 are the inertias of the link and the rotors (eﬀective, i.e. multiplied by the squared gear ratio); Ui,max , i = 1, 2 is the maximum torque for motor i; φi , i = 1, 2 represent the impedance of deformable elements as functions of the position of the rotors and link. A polynomial nonlinear stiﬀness model is used, whereby the applied force as a function of end-point displacement is φi (xj , xk ) = K1 (xj − xk ) + K2 (xj − xk )3 .

(2)

This model has been found to ﬁt well experimental data for the device in ﬁg. 2-a. The safety constraint HIC(x˙ mov , x˙ rot1 , x˙ rot2 , φ1 , φ2 ) ≤ HICmax describes the fact that the Head Injury Coeﬃcient of an hypothetical impact at any instant during motion, should be limited. The HIC is an empirical measure of biological damage used in car crash analysis literature ([12]), and depends on both the velocity of the impacting mass, its inertia and the eﬀective inertia of rotors reﬂected through the transmission stiﬀness. The HIC function is rather complex, and can only be evaluated numerically for non trivial cases. However, based on simulation studies, a conservative approximation of the HIC function for the antagonistic mechanisms was obtained which allows rewriting the safety constraint in the simpler form |x˙ mov | ≤ vsaf e (φ1 , φ2 , HICmax ) (cf. [2]). Diﬀerent solutions of problem (1) have been obtained numerically, setting parameters to realistic values as Mmov = 0.1 Kgm2 , Mrot1 = Mrot2 = 2.5 0.6 Kgm2 , HICmax = 100 ms4 , U1,max = U2,max = 7.5 Nm. A ﬁrst interesting set of results is reported in ﬁg. 3. The optimal proﬁles of link velocity and joint stiﬀness are reported for the case where both the initial and ﬁnal conﬁgurations are required to be stiﬀ (σ0 = σf = 16 Nm/rad, plots a and b) and when both are compliant (σ0 = σf = 0.2 Nm/rad, plots c and d). Notice in ﬁg. 3 that the stiﬀ-and-slow/fast-and-soft paradigm applies also to antagonistic actuation approaches. The minimum time necessary in the two cases is 2.4 sec and 2.65 sec, respectively. This level of performance should be compared with what can be achieved by a simpler actuation system, consisting of a single actuator connected to the link through a linear elastic element (this arrangement is sometimes referred to as SEA, Series Elastic Actuation [9]). A SEA system with a motor capable of torque Umax = 2U1,max = 15 Nm and inertia Mrot = 2Mrot1 = 1.2 Kgm2 , with linear elasticity coeﬃcient matched exactly with the required stiﬀness in the two cases σ = 16 and σ = 0.2, would

Optimization and Fail–Safety of Antagonistic Actuation for pHRI 4.5

16

4

14

3

Stiffness [Nm/rad]

Link Velocity [rad/s]

3.5

2.5 2 1.5

12 10 8 6 4

1

2

0.5 0 0

113

0.5

1 Time [s]

1.5

0 0

2

0.5

1 Time [s]

a)

1.5

2

b)

4.5

2.5

4 2

3

Stiffness [Nm/rad]

Link Velocity [rad/s]

3.5

2.5 2 1.5 1

1.5

1

0.5

0.5 0 0

0.5

1

1.5 Time [s]

c)

2

2.5

0 0

0.5

1

1.5 Time [s]

2

2.5

d)

Fig. 3. Optimization results for antagonistic actuation without cross-coupling in a pick-and-place task. A stiﬀ-to-stiﬀ task is shown in a, b, while c, d, refer to a soft-to-soft task.

reach the desired conﬁguration in 3.15 sec and 3.6 sec, respectively. Moreover, it should be pointed out that the SEA system cannot change its stiﬀness without modifying the mechanical hardware. Focusing again on the antagonistic VIA system’s results, it is also interesting to notice that, in the likely case that the task requires the manipulator to be stiﬀ at the initial and ﬁnal conﬁgurations (as it would happen e.g. in a precision pick-and-place task), the actuators are required to use a signiﬁcant portion of their maximum torque just to set such stiﬀness, by co–contracting the elastic elements. However, it is also in the initial and ﬁnal phases that torque should be made available for achieving fastest acceleration of the link. Based on this observation, it can be conjectured that some level of preloading of the nonlinear elastic elements in an antagonistic VIA system could be beneﬁcial to performance. Elastic cross-coupling between actuators (ﬁg. 2-b) can have a positive eﬀect in that it can bias the link stiﬀness at rest, so that more torque is available in slow phases, while torque is used for softening the link in fast motion. On the other hand, it is intuitive that very stiﬀ crosscoupling elements would drastically reduce the capability of the mechanism to vary link stiﬀness, thus imposing low velocities for safety and ultimately a performance loss. It is therefore interesting to study the eﬀect of cross-coupling, to determine if there is an intermediate value of stiﬀness which enhances performance with respect to the limit cases of ﬁg. 1 (no cross-coupling) and constant stiﬀness (rigid cross-coupling). To this purpose, we study a modiﬁed formulation of the safe brachistochrone problem, namely

114

G. Boccadamo et al. 5

16

3.6 3.4 3.2 3 2.8 2.6 2.4 2.2

4

Stiffness [Nm/rad]

Link Velocity [rad/s]

Minimum Time [s]

3.8

3 2 1

14 12 10 8 6 4 2

−2

10

0

2

10

10

K* [Nm/rad]

a)

0 0

0.5

1

1.5

Time [s]

b)

2

2.5

0 0

0.5

1

1.5

Time [s]

2

2.5

c)

Fig. 4. a) Minimum time to reach the target conﬁguration vs. cross-coupling stiﬀness K3 (L0 = 1). Solid: σ0 = σf = 16 Nm/rad; dashed: σ0 = σf = 0.2 Nm/rad. Optimization results for antagonistic actuation with cross-coupling in a pick-andplace, stiﬀ-to-stiﬀ task (solid) and soft-to-soft task (dashed): velocity (b) and joint stiﬀness (c).



T

minτ 0 1 dt Mrot1 x ¨rot1 + φ1 (xrot1 , xmov ) − φ3 (xrot1 , xrot2 ) = τ1  Mrot2 x ¨rot2 + φ2 (xrot2 , xmov ) + φ3 (xrot1 , xrot2 ) = τ2 Mmov x ¨link + φ2 (xrot2 , xmov ) − φ1 (xrot1 , xmov ) = 0 |τ1 | ≤ U1,max |τ | ≤ U2,max  2 HIC(x˙ mov , x˙ rot1 , x˙ rot2 , φ1 , φ2 ) ≤ HICmax ,

(3)

where φ3 (xrot1 , xrot2 ) indicates the cross-coupling elasticity. In particular, we study how optimal solutions vary in diﬀerent instances of the problem with increasing cross-coupling stiﬀness. A linear stiﬀness model is assumed for crosscoupling, with φ3 (xrot1 , xrot2 ) = K3 (L0 + xrot2 − xrot1 ), where L0 denotes the preload oﬀset. In ﬁg. 4 numerical solutions obtained for problem (3) are reported, for the two stiﬀ-to-stiﬀ and soft-to-soft tasks already considered in ﬁg. 3). It can be observed from ﬁg. 4-a that an optimal value of cross-coupling K3 exists in both cases, but they do not coincide. Setting for instance K3 = 0.04 Nm/rad, the optimal value for the stiﬀ-to-stiﬀ task, we obtain the optimal link velocity and stiﬀness plots reported in ﬁg. 4-b and -c, and the optimal times 2.35 sec and 2.65 sec, respectively. Optimizing the design for one task can hence decrease performance in others.

3 Fail Safety In this section we analyze how antagonistic VIA mechanisms behave under failure of some of their components, and compare their ability to remain safe in spite of failures, with those of other possible actuation structures. A large variety of possible failure modes should be considered in order to assess fail safety of a robot system. In our case, we focus on the following

Optimization and Fail–Safety of Antagonistic Actuation for pHRI

115

possible events: mechanical failures, consisting in breakage of one or more of the elastic elements, leading to K → 0, and HW/SW control failures, whereby one or more of the actuators delivers an uncontrolled torque. In the latter case, we consider that the torque can either remain “stuck” at some value, or default to zero, or to ±Umax . An extensive simulation campaign has been undertaken to study the eﬀective HIC at impacts of an antagonistic VIA mechanism (ﬁg. 1-b), evaluating all possible failure mode combinations. To be conservative, the HIC is evaluated assuming an impact occurs after some time from the instant of the fault event, during which time the mechanism could increase its momentum. The diagram in ﬁg. 5 reports the worst-case HIC during the execution of a typical pick-and-place task, under a nominal optimal control. Labels in different regions indicate which failure mode determines the worst-case HIC of impacts. It can be noticed from ﬁg. 5 that the worst case impact has maxi-

100 τ =U 1 1,max τ2 = U2,max

90 80

K =0 1 τ2 = U2,max

4

50

HIC [m .5/s ]

60

2

70

τ1 = U1,max τ2 = U2,max

τ1 = stuck τ2 = stuck

40

K1 = 0 τ2 = stuck

30 20 10 0

1

1.5

2

2.5 Time [S]

3

3.5

Fig. 5. HIC values under diﬀerent failures mode for VIA.

mum HIC value around 100. This has been obtained by suitably choosing the mechanism parameters, in particular by setting U1,max = U,2max = 1.5N m. The performance under these design choice is around 3 sec. Should diﬀerent values of worst-case HIC be speciﬁed, the antagonistic VIA mechanism could be designed with diﬀerent parameters to accommodate for the speciﬁcations. It is interesting to compare these results with other existing actuation systems for pHRI. Namely, we will make reference to SEA and to the so-called Distributed Macro-Mini (DM2 ) actuation scheme ([15]). The latter basically consists of a SEA with a small inertia, low torque, high bandwidth actuator connected directly to the link inertia to actively damp oscillations. In order to obtain a meaningful comparison, we consider optimized SEA and DM2 implementations, with equal rotor and link inertia, which obtain

116

G. Boccadamo et al.

a similar performance of 3 sec in the same task. Fig. 6 describes the worstcase HIC values obtained for SEA and DM2 , respectively. Failure modes have been considered which are analogous to those described above, as are all other parameters in the simulations. It can be observed that, in some modes (notably for control torques defaulting to their maximum value, τ → Umax ), both SEA and DM2 are not as safe as VIA (HIC 100). The DM2 scheme exhibits slightly better fail safety characteristics than SEA. An explanation of the apparently superior fail-safety characteristics of antagonistic VIA actuation is that such scheme achieves comparable nominal performance by employing two motors each of much smaller size than what necessary in the SEA and DM2 .

HIC (SEA) [m

2.5 4

/s ]

200 150

1

MAX

τ = UMAX

τ = UMAX τ = UMAX

50 0

τ = −U

τ1 = −UMAX

100

1

1.5

2

2.5 Time [s]

3

3.5

τ1 = −U1,MAX τ2 = τ2, opt

200 τ =U 1 1,MAX τ2 = U2, MAX

2

HIC (DM ) [m

2.5 4

/s ]

300

100 0

1

1.5

τ1 = −U1,MAX τ2 = τ2, opt τ =U 1 1,MAX τ =U 2

2

2.5 Time [s]

3

4

3.5

2, MAX

4

Fig. 6. Maximum of the HIC value in various failure cases for SEA (top) and DM2 (bottom).

4 Conclusion We discussed the design of variable stiﬀness actuation systems based on antagonistic arrangements of nonlinear elastic elements and motors. The basic stiﬀ-and-slow/fast-and-soft idea of the VIA approach was shown to apply and to be eﬀective for realistic models of antagonistic actuation for a single joint. However, a VIA mechanism assembly can sometimes become heavy for practical purposes to be implemented in a serial link robot. It may be advantageous to use the actuators at base instead of respective joints, and motion can be transferred to joints via transmissions, where, the transmission element itself

Optimization and Fail–Safety of Antagonistic Actuation for pHRI

117

can be designed to have nonlinear stiﬀness, for example in a tendon driven system. In order for the system to guarantee safety of operations in the proximity of humans, its behaviour must remain safe in conditions where functionality is degraded by possible failures. A thorough examination of possible HW and SW failures of the system has been considered, and results have been compared with those of other possible solutions, with favourable result.

Acknowledgments Work partially done with the support of the EU Network of Excellence EURON - European Robotic Network, under the Prospective Research Project PHRIDOM

References 1. A. Albu-Schaﬀer, A. Bicchi, G. Boccadamo, R. Chatila, A. De Luca, A. De Santis, G. Giralt, G. Hirzinger, V. Lippiello, R. Mattone, R. Schiavi, B. Siciliano, G. Tonietti, and L. Villani. Physical human-robot interaction in anthropic domains: Safety and dependability. 4th IARP/IEEE-EURON Workshop on Technical Challenges for Dependable Robots in Human Environments, 2005. 2. A. Bicchi and G. Tonietti. Fast and soft arm tactics: Dealing with the safety– performance tradeoﬀ in robot arms design and control. IEEE Robotics and Automation Magazine, Vol. 11, No. 2, June, 2004. 3. D.G. Caldwell, G.A. Medrano-Cerda, and M.J. Goodwin. Braided pneumatic actuator control of a multi-jointed manipulator. In Systems Engineering in Service of Humans, Proceedings International Conference on, IEEE Systems Man and Cybernetics, volume 1, pages 423–428, 1993. 4. G. Giralt and P. Corke, editors. Technical Challenge for Dependable Robots in Human Environments, Seoul, Korea, 2001. IARP/IEEE Workshop. 5. J. Guiochet and A. Vilchis. Safety analysis of a medical robot for teleechography. In Second IARP IEEE/RAS joint workshop on Technical Challenge for Dependable Robots in Human Environments, Toulouse, France, October 2002. LAAS-CNRS. 6. J. Heinzmann and A. Zelinsky. The safe control of human–friendly robots. In Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pages 1020–1025, 1999. 7. K. Ikuta, H. Ishii, and M. Nokata. Safety evaluation method of design and control for human-care robots. The International Journal of Robotics Research, 22(5):281–297, May 2003. 8. I. Mizuuchi, R. Tajima, T. Yoshikai, D. Sato, K. Nagashima, M. Inaba, Y. Kuniyoshi, and H. Inoue. Design and control of the ﬂexible spine of a fully tendondriven humanoid ”kenta”. In Proceedings of the 2002 IEEE/RSJ Intl. Conference on Intelligent RObots and Systems, EPFL, Lausanne, Switzerland, pages 2527–2532, 2002.

118

G. Boccadamo et al.

9. G.A. Pratt and M. Williamson. Series elastics actuators. In Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pages 399–406, 1995. 10. G. Tonietti, R. Schiavi, and A. Bicchi. Design and control of a variable stiﬀness actuator for safe and fast physical human/robot interaction. In Proc. ICRA, pages 528–533, 2005. 11. R.W. van der Linde. Design, analysis, and control of a low power joint for walking robots, by phasic activation of mckibben muscles. IEEE Transactions on Robotics and Automation, 15(4):599–604, August 1999. 12. J. Versace. A review of the severity index. In Proc. of the Fifteenth Stapp Car Crash Conference, number SAE Paper No. 710881, pages 771–796. Society of Automotive Engineers, 1971. 13. J. Yamaguchi, S. Inoue, D. Nishino, and A Takanishi. Development of a bipedal humanoid robot having antagonistic driven joints and three dof trunk. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, Victoria, B.C., Canada, pages 96 – 101, 1998. 14. J. Yamaguchi, D. Nishino, and A Takanishi. Realization of dynamic biped walking varying joint stiﬀness using antagonistic driven joints. In Proceedings of the IEEE International Conference on Robotics and Automation, Leuven, Belgium, pages 2022–2029, 1998. 15. M. Zinn, O. Khatib, B. Roth, and J.K. Salisbury. A new actuation approach for human friendly robot design. In Proc. of Int. Symp. on Experimental Robotics - ISER’02, 2002.

Bilateral Control of Diﬀerent Order Teleoperators Jos´e M. Azor´ın1 , Rafael Aracil2 , Jos´e M. Sabater1 , Manuel Ferre2 , Nicol´ as M. Garc´ıa1 and Carlos P´erez1 1 2

Universidad Miguel Hern´ andez de Elche, Avda. de la Universidad s/n, 03202 Elche (Alicante), Spain [email protected] DISAM, ETSII, Universidad Polit´ecnica de Madrid, 28006 Madrid, Spain [email protected]

This paper presents a bilateral control method for teleoperation systems where the master and the slave are modeled by diﬀerent order transfer functions. The proposed methodology represents the teleoperation system on the state space and it is based in the state convergence between the master and the slave. The method allows that the slave follows the master, and it is able to establish the dynamic behavior of the teleoperation system. The ﬁrst results obtained when the method is being applied to a commercial teleoperation system in which the master and the slave are modeled by diﬀerent order discrete transfer functions are shown in this paper.

1 Introduction Often, in a teleoperated system, the interaction force of the slave with the environment is reﬂected to the operator to improve the task performance. In this case, the teleoperator is bilaterally controlled [1]. The classical bilateral control architectures are the position–position [2] and force–position architecture [3]. Additional control schemes have been proposed in the literature, e.g. the bilateral control for ideal kinesthetic coupling [4] or the bilateral control based in passivity to overcome the time delay problem [5]. Usually the proposed bilateral control schemes consider simple master and slave models of the same order. They do not provide a design procedure to calculate the control system gains when the master and the slave are modeled by diﬀerent order transfer functions. In [6] we presented a bilateral control method of teleoperation systems with communication time delay. In this method, the teleoperation system is modeled on the continuous-time state space from nth-order linear diﬀerential equations that represent the master and the slave. Through the state convergence between the master and the slave, the control method achieves that

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 119–127, 2006. © Springer-Verlag Berlin Heidelberg 2006

120

J.M. Azor´ın et al.

the slave follows the master and it allows to establish the desired dynamics of the teleoperation system. However, this control method considers only teleoperation systems where the master and the slave are modeled by diﬀerential equations of the same order. This paper explains how the state convergence methodology can be used to control teleoperation systems where the master and the slave are modeled by diﬀerent order discrete transfer functions. In addition, in this paper, diﬀerent from [6], the teleoperation system is modeled in the discrete-time domain, and the communication time delay is not considered in order to simplify the explanation. Clearly, the results presented in this paper could be directly applied to continuous-time teleoperation systems with communication time delay. The paper is organized as follows. Section 2 describes the bilateral control methodology of teleoperation systems based in the state convergence. The application of this methodology to control diﬀerent order teleoperation systems is explained in section 3. Section 4 shows the results obtained when a master and an slave with diﬀerent order transfer functions have been controlled using the method proposed in the paper. Finally, section 5 summarizes the conclusions of this paper.

2 Bilateral Control by State Convergence This section describes the bilateral control method of teleoperation systems based in the state convergence [6]. However, diﬀerent from [6], a teleoperation system without communication time delay where the master and the slave are modeled by nth-order discrete transfer functions is considered. The control method is based on the state space formulation and it allows that the slave follows the master through state convergence. The method is able also to establish the desired dynamics of this convergence and the dynamics of the slave manipulator. 2.1 Modeling of the Teleoperation System The modeling of the teleoperation system is based on the state space formulation, Fig. 1. The master system is represented in the state space like: xm (k + 1) = Gm xm (k) + Hm um (k) ym (k) = Cm xm (k)

(1)

and the slave system is represented similarly. The matrices that appear in the model are: G2 , inﬂuence in the slave of the operator force; Km and Ks , feedback matrices of the master and slave state; and Rm and Rs , interaction slave - master, and interaction master slave.

Bilateral Control of Diﬀerent Order Teleoperators Km

Fm

um

+ +

Hm

+

MASTER

+

xm

-1

z

+

Rm

G2

121

Cm

ym

Gm

Rs

SLAVE us

+ +

Hs

+

+

-1

+

z

xs

Cs

ys

Gs

Ks

Fig. 1. Modeling of the teleoperation system

From the model shown in Fig. 1, the next state equation of the teleoperation system is obtained: xs (k) H s G2 Fm (k) + xm (k) Hm (2) where Fm represents the force that the operator applies to the master. The master and the slave are modeled by nth-order discrete transfer functions. The representation of the master is given by: xs (k + 1) Gs + Hs K s Hs R s = xm (k + 1) Hm R m Gm + H m K m

Gm (z) =

zn

bmn−1 z n−1 + · · · + bm1 z + bm0 + amn−1 z n−1 + · · · + am1 z + am0

(3)

and the representation of the slave is analogous. The master and the slave are modeled in the state space using the controller canonical form. Considering this master and slave represention, the matrices Km , Ks , Rm and Rs are row vectors of n components, and G2 is a real number. If the environment is modeled by means of a stiﬀness (ke ), the matrix Ks can be used to incorporate in the teleoperation system model the interaction

122

J.M. Azor´ın et al.

of the slave with the environment. In the same way, the matrix Rm can be used to consider force feedback from the slave to the master assuming a force feedback gain kf . 2.2 Control Method Through State Convergence In the teleoperation system model shown in Fig. 1, there are 3n + 1 control gains that are necessary to obtain: Km (n components), Ks (n components), Rs (n components) and G2 (one component). To calculate these control gains it is necessary to get 3n + 1 design equations. Applying the next linear transformation to the system (2), the error state equation between the master and the slave is obtained: xs (k) I 0 = Es xs (k) − Em xm (k) Es −Em

xs (k) xm (k)

(4)

where Es = diag{bs0 , · · · , bsn−1 } and Em is similar to Es . Let xe (k) be the error between the slave and the master, xe (k) = Es xs (k)− Em xm (k). If the error converges to zero, the slave output will follow the master output. From the error state equation, n + 1 design equations can be obtained to achieve that the error evolves as an autonomous system, and the slave output follows the master output. In addition, n − 1 conditions between the numerator coeﬃcients of the master and slave discrete transfer functions can be derived. If these conditions are not satisﬁed, there will be an error between the master and slave output. When the error evolves like an autonomous system, the characteristic polynomial of the system can be calculated. From this polynomial, 2n design equations can be obtained to establish the slave and the slave-master error dynamics. These 2n equations plus the n + 1 previous equations, form a system of 3n + 1 equations. Solving this equations system, it will be possible to obtain the 3n + 1 control gains. With these control gains, the slave manipulator will follow the master and the dynamics of the error and the slave will be established.

3 Application to Teleoperation Systems of Diﬀerent Order The control methodology that considers teleoperation systems where the master and the slave are modeled by discrete transfer functions (DTFs) of the same order has been presented in the previous section. This section explains how to use the design equations to control teleoperation systems where the master and the slave are modeled by diﬀerent order DTFs. If the master and the slave are modeled by DTFs of the same order the control method allows that the slave follows the master, and it is able also

Bilateral Control of Diﬀerent Order Teleoperators

123

to establish the dynamics of the slave and the slave-master error (i.e. to ﬁx the n poles of the slave and the n poles of the error). If the master and the slave are modeled by diﬀerent order DTFs, there are two options to design the control system: to increase the order of the smaller DTF, or to reduce the order of the higher DTF. Next, each option is explained describing its eﬀects in the control system. 3.1 To Increase the Order of the Smaller DTF In this option the order of the smaller DTF is increased to achieve that it has the same order of the higher DTF. The order is increased adding the necessary pole-zero pairs. Pole-zero pairs are added to avoid the increment of the delay attached to the system. The pole is placed in such a way that it does not aﬀect to the system dynamics. The zero is placed near the pole, but avoiding the exact cancellation, in order to increase really the DTF order. It must be checked that the increased DTF has the same static gain that the original DTF. Below the eﬀects of increasing the order of the smaller DTF in the control are described. First, if a zero is added, the conditions to achieve the evolution of the error as an autonomous system could not be veriﬁed, and a constant error between the master and the slave could exist. This is not a problem, because, e.g., in a teleoperation system where the master dimensions are smaller than the slave dimensions, and the output of both was the cartesian position, it will not be desirable that both positions were the same. The eﬀects of increasing the master order are: • The slave dynamics is completely established. • All the desired error poles are ﬁxed. However, because of the fact that the error has the same number of poles that the slave order, and the master order is smaller, part of the error depending on the master would be artiﬁcially established. On the other hand, the eﬀects of increasing the slave order are: • Additional poles of the slave dynamics are established because the slave order has been increased. This is not a problem because all the poles of the original slave dynamics are ﬁxed. • All the desired error poles are established. However, in a similar way to the previous case, part of the error depending on the slave would be artiﬁcially established 3.2 To Reduce the Order of the Higher DTF In this option the order of the higher DTF is reduced to achieve that it has the same order of the smaller DTF. That is, the same number of pole-zero pairs that there are in the smaller DTF will be considered to design the control

124

J.M. Azor´ın et al.

system. Therefore some pole-zero pairs of the higher DTF will be removed in the design. In order to remove a pole-zero pair it would be desirable to select the pole further from the dominant poles, and to select a zero near the pole to consider a pole-zero cancellation. Below the eﬀects of reducing the order of the higher DTF in the control are described. First, if a dominant pole is removed, or a zero that is not near the pole selected is removed, the reduced DTF will be very diﬀerent to the original DTF and the control gains obtained could not be applied to the real teleoperation system. In addition, when a zero is removed, the conditions to achieve the evolution of the error as autonomous system could not be veriﬁed. The eﬀects of reducing the slave order are: • The slave dynamics is not completely established, because fewer poles have been considered in the design phase. • Only the number of error poles ﬁxed by the master order can be established. Therefore, part of the error depending on the slave will not be established. On the other hand, the eﬀects of reducing the master order are: • The slave dynamics is completely established. • Only the number of error poles ﬁxed by the slave order can be established. Therefore, part of the error depending on the master will not be established. 3.3 Comparison of Options If the order of the smaller DTF is increased, the control method can completely establish the dynamics of the teleoperation system, i.e. the slave and the error dynamics is ﬁxed. On the other hand, if the order of the higher DTF is reduced, the dynamics of the teleoperation system can not be completely established. In this case, part of the error dynamics depending on the reduced DTF can not be ﬁxed. In addition, if the slave order is reduced, the slave dynamics will not be completely established. In both options, when the DTF order is increased or reduced, the conditions to achieve the evolution of the error as autonomous system could not be veriﬁed, but this is not a problem. Therefore the best option to design the control system of a teleoperator where the master and the slave have diﬀerent order is increasing the order of the smaller DTF. The control gains that are obtained modifying the master or the slave DTF in the design phase must be used to control the real teleoperation system. The state convergence methodology allows that the control gains can be directly applied to the real teleoperation system because of the robustness of the control method to the uncertainty of the design parameters [7]. On the other hand, state observers must be designed to apply the control. In the case of the reduced DTF, the state observer must estimate the number of state variables ﬁxed by the higher DTF, so it must be designed using the increased DTF.

Bilateral Control of Diﬀerent Order Teleoperators

125

4 Experimental Results The EL (Elbow Pivot) joint of the Grips teleoperation system has been bilaterally controlled in simulation mode using the state convergence methodology. The Grips manipulator system from Kraft Telerobotics is a six DOF teleoperator with force-feedback. The slave is an hydraulic manipulator, and the master is powered by electrical motors. The identiﬁed DTFs for the EL joint of the master and the slave, considering a sample time of T = 0.0005s, are, respectively: 3.79 × 10−7 z 2 (5) Gm (z) = 3 z − 2.974z 2 + 2.948z − 0.9741 1.357 × 10−5 z 1.357 × 10−5 z Gs (z) = 2 = (6) z − 1.9928z + 0.9928 (z − 1)(z − 0.9928) The process of identiﬁcation has been done using the corresponding Matlab Toolbox. Both DTFs have been identiﬁed considering the Box-Jenkins model. As it has been previously explained, when the order of the master and slave DTF is diﬀerent, the best option to design the control system is increasing the order of the smaller DTF. Therefore the order of the slave DTF must be increased. A pole-zero pair has been added in the slave DTF. The pole has been added far from the dominant poles to avoid that it aﬀects to the system dynamics. And the zero has been placed near the pole: 1.357 × 10−5 z 2 − 1.262 × 10−5 z 1.357 × 10−5 z(z − 0.93) = 3 (z − 1)(z − 0.9928)(z − 0.9302) z − 2.923z 2 + 2.8465z − 0.9235 (7) In order to solve the design equations of the control method, all the numerator coeﬃcients of (5) and (7) must not be null. In this case, the coeﬃcients {bm1 , bm0 , bs0 } are null, i.e. there are some zeros placed in z = 0. For this reason, the zeros placed in z = 0 have been slightly modiﬁed, and the DTFs considered for the design are the following: Gs (z) =

Gm (z) =

3.79 × 10−7 z 2 + 10−8 z + 10−8 z 3 − 2.974z 2 + 2.948z − 0.9741

(8)

1.357 × 10−5 z 2 − 1.262 × 10−5 z + 10−8 (9) z 3 − 2.923z 2 + 2.8465z − 0.9235 In addition, to verify the conditions of the control method between the numerator coeﬃcients of Gm (z) and Gs (z), and to achieve the evolution of the error as autonomous system, the numerator coeﬃcients of Gs (z) in (9) have been modiﬁed for the design in this way: Gs (z) =

Gs (z) =

1.357 × 10−5 z 2 + 3.58 × 10−7 z + 3.58 × 10−7 z 3 − 2.923z 2 + 2.8465z − 0.9235

(10)

The control gains have been obtained considering that the force feedback gain is kf = 0.1, the slave interacts with a soft environment (ke = 10N m/rad), and that the poles of the error and the slave are placed in z = 0.95.

126

J.M. Azor´ın et al.

If the control gains are applied to the teleoperation system considered for the design, DTF (8) and (10), the slave follows the master without error, left part of Fig. 2. However, if they are applied to the increased original system, DTF (5) and (7), the slave follows the master, but there is a constant error because the conditions to achieve the evolution of the error as autonomous system are not veriﬁed, right part of Fig. 2. In both cases, a unitary constant operator force has been considered. −3

3.5

−3

x 10

3.5 3

master slave

2.5

master and slave position

master and slave position

3

2 1.5 1 0.5 0

x 10

master slave

2.5 2 1.5 1 0.5

0

0.5

1 time (s)

1.5

2

0

0

0.5

1 time (s)

1.5

2

Fig. 2. Master and slave position considering DTF (8) and (10) (left part), and DTF (5) and (7) (right part)

In order to verify the performance of the control over the original teleoperation system, two state observers have been designed for the identiﬁed DTFs of the EL joint (DTF (5) and (6)). Both state observers must estimate three state variables. The design of the state observer for the DTF (5) is straight. However, as the order of the DTF (6) is two, the observer has been designed from the DTF (7). In the left part of Fig. 3 the position evolution of the original teleoperation system using the designed observers is shown. It can be veriﬁed that the results obtained are the same as the shown in the right part of Fig. 2. The control signals applied to the master and the slave are shown in the right part of Fig. 3. In this ﬁgure, the slave control signal has initial values close to zero, but not null. These results allow verify that the control method by state convergence can be applied to teleoperation systems where the master and the slave are modeled by diﬀerent order DTFs.

5 Conclusions This paper has presented a bilateral control method for teleoperation systems where the master and the slave are modeled by diﬀerent order transfer functions. It has been veriﬁed that the order of the smaller DTF must be increased

Bilateral Control of Diﬀerent Order Teleoperators

127

−3

3.5

x 10

1

master slave

master and slave control signal

3

master and slave position

master slave

0.8

2.5 2 1.5 1

0.6 0.4 0.2 0 −0.2 −0.4 −0.6

0.5

−0.8 0

0

0.2

0.4 0.6 time (s)

0.8

1

0

0.2

0.4 0.6 time (s)

0.8

1

Fig. 3. Master and slave position (left part), and control signal (right part)

to design the control system. Then, state observers that estimate the number of state variables ﬁxed by the higher DTF must be designed in order to apply the control. The control method allows that the slave follows the master, and it is able to establish the dynamic behavior of the teleoperation system. In the paper the performance of the control scheme has been tested in simulation mode over the EL joint of the Grips teleoperation system. The next work is controlling the real joint. The ﬁnal aim is controlling all the joints of the system applying the state convergence methodology.

References 1. B. Hannaford. Stability and performace tradeoﬀs in bilateral telemanipulation. In Proc. IEEE Int. Conf. on Robotics and Automation, pages 1764–1767, 1989. 2. R.C. Goertz and al. The anl model 3 master-slave manupulator design and use in a cave. In Proc. 9th Conf. Hot Lab. Equip., volume 121, 1961. 3. C.R. Flatau. Sm 229, a new compact servo master-slave manipulator. In Proc. 25th Remote Syst. Tech. Div. Conf., volume 169, 1977. 4. Y. Yokokohji and T. Yoshikawa. Bilateral control of master-slave manipulators for ideal kinesthetic couplig–formulation and experiment. IEEE Transactions on Robotics and Automation, 10 (5):605–620, 1994. 5. R.J. Anderson and M.W. Spong. Bilateral control for teleoperators with time delay. IEEE Transactions on Automatic Control, 34 (5):494–501, 1989. 6. J. M. Azorin, O. Reinoso, R. Aracil, and M. Ferre. Generalized control method by state convergence of teleoperation systems with time delay. Automatica, 40 (9):1575–1582, 2004. 7. J. M. Azorin, O. Reinoso, R. Aracil, and M. Ferre. Control of teleoperators with communication time delay through state convergence. Journal of Robotic Systems, 21 (4):167–182, April 2004.

Navigation and Planning in an Unknown Environment Using Vision and a Cognitive Map Nicolas Cuperlier, Mathias Quoy, and Philippe Gaussier ETIS-UMR 8051 Universit de Cergy-Pontoise - ENSEA 6, Avenue du Ponceau 95014 Cergy-Pontoise France [email protected] Summary. We present a framework for Simultaneous Localization and Map building of an unknown environment based on vision and dead-reckoning systems. An omnidirectional camera gives a panoramic image from which no a priori deﬁned landmarks are extracted. The set of landmarks and their azimuth relative to the north given by a compass deﬁnes a particular location without any need of an external environment map. Transitions between two locations are explicitly coded. They are simultaneously used in two layers of our architecture. First to construct, during exploration (latent learning), a graph (our cognitive map) of the environment where the links are reinforced when the path is used. And second, to be associated, on an another layer, with the integrated movement used for going from one place to the other. During the planning phase, the activity of transition coding for the required goal in the cognitive map spreads along the arcs of this graph giving transitions (nodes) an higher value to the ones closer from this goal. We will show that, when planning to reach a goal in this environment is needed, the interactions of these two levels can lead to the selection of multiple transitions corresponding to the most activated ones according to the current place. Those proposed transitions are ﬁnally exploited by a dynamical system (neural ﬁeld) merging these informations. Stable solution of this system gives a unique movement vector to apply. Experimental results underline the interest of such a soft competition of transition information over a strict one to get a more accurate generalization on the movement selection.

1 Introduction Path planning requires from the agent or the robot to select the appropriate action to perform. This task might be complex when several actions are possible, and so diﬀerent approaches have been proposed to choose what to do next. Experiments carried out on rats have led to the deﬁnition of cognitive maps used for path planning [21]. Most of cognitive maps models are based on graphs showing how to go from one place to an other [2, 3]. They mainly diﬀer in the way they use the map in order to ﬁnd the shortest path, in the way

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 129–142, 2006. © Springer-Verlag Berlin Heidelberg 2006

130

N. Cuperlier, M. Quoy, and P. Gaussier

they react to dynamical environment changes, and in the way they achieve contradictory goal satisfactions. One can refer to [9, 12] for a comparative review of localisation and mapping models. Many methods rely on the combination of diﬀerent algorithms that have to be triggered appropriately (and concurrently) when necessary. For instance, localisation may involve diﬀerent sensors (laser, ultra-sound, visual feature recognition ...) that have to be chosen appropriately. Some works use ruled-based algorithms, classical functional approach, that can exhibit the desired behaviors, we will not discuss them in this paper, but one can refer to [8]. Instead, other works try to look at what the nature does by taking inspiration from neurobiology to design control architectures. There are at least two reasons for this: • ﬁrst, getting robust, adaptive, opportunistic and ready-made solutions for control architecture. • second, if robotic results can be compared to experimental results involving several parts of the brain, which are generally diﬃcult to study due to its complexity, it can help neurobiologists to understand how a neurobiological model behaves. Hence we propose here a uniﬁed neuronal framework based on an hippocampal and and prefrontal model where vision, place recognition and deadreckoning are fully integrated (see Fig. 4 for an overview of the architecture). This model relies on a topological map: the environment is coded via a set of distinctive nodes and by the way a robot can go from one node to another. In our work, those nodes are not directly places of the environment but rather the transition between two of them. No cartesian metric informations and no occupancy grid are used to construct the map. Localisation is achieved using a biomimetic model designed to emulate the neural activity of particular neurons found in the rat hippocampus named place cells. Those cells learn direction and identity (recognition) of punctual landmarks leading to a place deﬁnition. A key point for the understanding of this model is the distinction between transition coding for the succession of two place cells at the recognition (visual) level and motor transition encoding, on a motor level, the integrated movement performed to go from one place to the other. Whereas these two kinds of transition are strongly dependent and linked in a unique way, they do not have the same modality: one is related to vision coding and the other is related to motor coding. Keeping this basic distinction in mind, we can list the assets of this model: • autonomous landmarks extraction based on characteristic points (section 2) • autonomous place building via place-cells-like neurons: there are no a priori predeﬁned squares, or world model (section 3) • autonomous creation of transitions at recognition level (recognition transitions). A neuron codes the succession of two recognized place cells without any combinatorial explosion (section 4).

Navigation and Planning in an Unknown Environment

131

• autonomous cognitive map building based on those recognition transitions between places, giving topological informations like adjacency of two transitions linked (i.e: if transitions AB and BC are linked: adjacency of the destination place cell AB with the place cell of origine of transition cell BC). But, this map can also give a kind of metric via the value of the arc of this graph. (section 5). • autonomous association of recognition transitions with integrated movement giving motor transitions which can be used for planning (section 6). • autonomous planning using both the cognitive map (graph of recognition transitions) and the corresponding motor transitions (section 7) • stable movement given by the stable ﬁxed point solution of a dynamical system (section 8) Drawbacks will be left for conclusion.

2 Autonomous Landmark Extraction Based on Characteristic Points Images are taken by a panoramic camera at a low resolution. This allows to handle lighter images that may be processed in a real time. In order to eliminate problems induced by luminance variability, we only use the gradient image as input of the system (a 1500 × 240 pixels image extracted from the 640 × 480 pixels panoramic image which is originally circular). Two processes then occur in parallel: • First, curvature points are extracted from this gradient image by Diﬀerence Of Gaussian (D.O.G) ﬁltering. Those curvature points are robust focal points due to the low resolution. Each focal point is the center of a 32x32 pixels small image giving the local visual area extracted around it (cf. circles on Fig. 1) . This image is binarized through a log-polar transform [19, 11] and next it is learned on a neuron coding for this landmark. A soft competition between landmark neurons, allowing several interpretations of a given local snapshot, is then computed • Second, each landmark is linked with its angular position relative to the north given by a compass [20, 14]. In a panoramic image, 32 (landmark, azimuth) pairs are extracted (see Fig. 1). Thus, this visual system provides both a what and a where informations: the recognition of a 32 × 32 pixels snapshot in log-polar coordinates, and the azimuth of the focal point. What and where informations are then merged in a product space that memorizes the incoming inputs during a given time. The number of landmarks needed is a balance between the robustness of the algorithm and the speed of the process. If all landmarks were fully recognized, only three of them will be needed. But as some of them may not be recognized in case of changing conditions (luminance; occlusion), taking a greater number

132

N. Cuperlier, M. Quoy, and P. Gaussier

Fig. 1. Image taken from a panoramic camera. Below are 15 examples of 32 × 32 log-polar transforms taken as landmarks and their corresponding position in the image.

is enough to guarantee the robustness. Moreover the log-polar transform gives some rotation and depth robustness.

3 Autonomous Place Building Each set of (landmark, azimuth) pairs, merged in the product space, is learned and thus characterizes one location. The neuron coding for this location is called a “place cell” (PC) as the one found in the rat’s hippocampus [14]. Place cell’s activity is the result of a matching function computing the distance between the learned set and the current set (the distance is computed only on the recruited neurons). Thus, the activity of the k th PC can be expressed as follows: Pk = N

1 lk

NL

L − θi ) ωik .fs (Li ).gd (θik

(1)

i=1

L with lk = i=1 ωik the number of landmarks used for the k th PC, where ωik = {0, 1} expresses the fact that landmark i has been used to encode PC k, with NL the number of learned landmarks, Li the activity of the landmark i, fs (x) the activation function of the neurons in the landmark recognition L the learned azimuth of the ith landmark for the k th PC, θi the group, θik azimuth of the current local view interpreted as the landmark i. d is the angular diﬀusion parameter which deﬁnes the shape of the function gd (x). The purpose of fs (x) and gd (x) is to adapt respectively the dynamics of what and where groups of neurons. They are deﬁned as follow :

Navigation and Planning in an Unknown Environment

gd (x) = 1 − fs (x) =

1 1−s

|x| d.π

133

+

[x − s]

+

where [x]+ = x if x > 0 , and 0 otherwise. The s parameter rescales the activity of the landmark neuron over s between 0 and 1. The d parameter modulates the weight of the angular displacement. More information on the neural model for place cell coding may be found in [10, 7].

An

R.T

Outdistance

Fig. 2. Idealised shape of a place cell (An ) response, according to the distance to the exact location where it has been learned. This cell has its maximal response in this place.

Recruitment of a new place cell for encoding a new location is performed autonomously, without any external signal. If the activities of all previously learned place cells are below a given recognition threshold (R.T), then a new neuron is recruited for coding this recently discovered location (see Fig. 2). Hence, the density of location learned depends on the level of this threshold. But it depends also on the position of the robot in the environment. Namely, more locations are learned near walls or doors because a fast change in the angular position of near landmarks, or in the (dis)appearance of landmarks may occur. In other locations, small changes produce a small variation in the place cell activity (see Fig. 2) . If at a given place, several cells responds with an activity greater than the R.T, a competition takes place so that the most activated cell wins and codes this location. These cells are created during the exploration of the unknown environment. We use random exploration but naturally other kind of algorithms can be used. At the end of this task, the environment is fully covered by place cells, so that in any part of it a place cell responds speciﬁcally for it (see Fig. 3.).

134

N. Cuperlier, M. Quoy, and P. Gaussier

Fig. 3. A simulated environment fully explored. Each region represents the response domain for which a particular place cell win the competition for recognition. After a full exploration the entire environment is covered by the place cell population.

4 Autonomous Building of Transitions at Recognition Level Two successively reached places are coded by a transition cell (see Fig. 4). Hence two successively reached places (A, then B) are coded by a transition cell (AB). A relevant question is about the growth of the number of these cells. Before to show the experimental results, we already can make two important remarques about this growth. • First, this growth is intimately linked with the number of place cells. This last one mainly depends on two parameters: – the value of R.T: the higher R.T is the bigger is the number of cell created. – the complexity of the environment: the number and the location of its landmarks and the number of obstacles found inside. • Second, we do not create all possible transitions but only physically feasible transitions between “adjacent” place cells. And since the number of a place cell neighbours is necessary limited (see Fig 3), the number of transition created is also limited. Hence, we have studied the ratio between created transition cells over created place cells for three environments of increasing complexity according to their obstacle conﬁguration (a single, a two and a four room environment). The tests have been performed in simulation using a virtual robot (or animat). For each simulation, 10 animats have explored their environment for 50000 cycles. This number has been chosen high enough to be sure that the animat has learned a complete cognitive map. The results shown here are the average on these 10 animat results. The ratio remains stable around the mean value 5.45 for all environments (see table 1.).

Navigation and Planning in an Unknown Environment

Env / RT nbp nbt ratio nbp nbt ratio nbp nbt ratio

0.97 133.8(2.85) 735.8(19.80) 5.49(0.06) 606.2(6.89) 3389.2(56.38) 5.59(0.08) 643.7(9,88) 3281,2(48,80) 5.09(0,04)

135

Table 1. Ratio of the number of place cells (nbp) created over the number of transitions created (nbt) according to the number of room in the environment: with one room (top line), with two rooms (middle line) and four rooms (bottom line). Standard deviation is given into brackets. This ratio remains stable. There are ﬁve times more transition cells than place cells.

Indeed, only a few transitions can be created from a given place cell, since a transition is a link between “adjacent” place cells and since the number of a place cell neighbours is necessary limited. So there is no combinatorial explosion on the number of created transitions and we do not need a ”full matrix” to create the transitions: a matrix of only M*N is enough. With M the number of place cells, and N the maximal number of possible neighbours.

Place recognition t−1

Azimuth

...

Landmark − azimuth Place recognition t

...

... ... Landmark

Place recognition t−1 Cognitive map Transition map

...

...

Recognition transitions Place recognition t

Motor transitions

...

...

Motor command

One to one links − No learning One to all links − Learning

Fig. 4. Sketch of the model. From left to the right: merging landmarks and their azimuth in a product space, then learning of the corresponding set on a place cell. Two successive place cells (the one at time t and the previous one at t − 1) deﬁne a transition cell. They are used to build up the cognitive map and are also linked with the corresponding motor transition that integrated the movement performed.

136

N. Cuperlier, M. Quoy, and P. Gaussier

5 Autonomous Cognitive Map Building Since our robotic model is inspired from the animat approach [13] we use three contradictory animal like motivations (eating, drinking, and resting). Each one is associated with a satisfaction level that decreases over time and increases when the robot is on the proper source. When a level of satisfaction falls bellow a given threshold, the corresponding motivation is triggered so that the robot has to reach a place allowing to satisfy this need. Hence this place becomes the goal to reach. More sources can be added and one can increase the number of sources associated with a given motivation. Each time a new transition is created, a new node is recruited in the cognitive map. This node is then linked with the previous transition used. When a transition leads to a place cell where a source can be found, a link between the corresponding motivation and the most active node on the map is created and set to one (latent learning), otherwise this link is set to zero. After some time, exploring the environment leads to the creation of the cognitive map (see Fig. 5). This map may be seen as a graph where each node is a transition and the arcs the fact that the path between these two transitions was used. We can give a ﬁxed value (lower than one) to each link at the creation time. This value is increased if the link is used, and decreased if it is not. After some time passed in the environment, some links are reinforced. These links correspond to paths that are often used. In particular, this is the case when some particular locations have to be reached more often than others (see section 7) [15].

Fig. 5. Cognitive map (in red) build by exploration of the environment. The triangles give the successive robot positions starting from the right to the goal (on the left). Landmarks are represented by blue crosses. The blue rectangle are obstacles.

Navigation and Planning in an Unknown Environment

137

6 Autonomous Motor Transitions Creation Each of these cells is linked with the direction used to go from the starting location to the ending location. For instance, going from place A to place B creates a recognition transition cell AB and the corresponding node on the map. In the same time another transition cell is created on the motor level. This motor transition associates the direction (relative to the north) for going from A to B with the node AB on the map. This direction integrates all direction changes performed between A and the creation of B using robot wheel encoders to compute elementary displacement vectors. Direction changes are induced by a new movement vector generated by the exploration mechanism (random exploration) or due to the obstacle avoidance mechanism. Hence the integrated vector (I.V) takes care of all these movement changes. Each time this transition is performed, the I.V. is averaged with the corresponding learned direction. The I.V is then reset when entering a diﬀerent place cell. The norm of this vector is also computed in the same way.

7 Autonomous Planning Using the Cognitive Map and Motor Transitions When a goal has to be reached, the transitions leading to it are activated via the links learned during exploration between those transitions and the corresponding motivation. This activation is then diﬀused on the cognitive map graph, each node taking the maximal incoming value which is the product between the weight on the link (lower than one) and the activity of the node sending the link. After stabilization, this diﬀusion process gives the shortest path between all nodes and the goal nodes. This is a neural version of the Bellman-Ford algorithm [5, 17] (see Fig.6). When the robot is in a particular location A, all possible transitions beginning with A are possible. The top-down eﬀect of the cognitive map is to bias the possible transitions such that the ones chosen by the cognitive map have a higher value. This value reﬂects a topological distance measure: the number of intermediate node to get in touch with to actually reach the goal. This small bias is enough to select/ﬁlter the appropriate transitions via a competition mechanism (however see section 8). This mechanism realizes a soft competition: several motor transitions are proposed at this level. They allow to compute a more accurate direction than a strict competition since transitions starting only from places really close to the current one and topologically close to the goal (on the graph) are selected. Final selection of the motor action results from the merging of these global decisions with local constraints such as obstacle avoidance, robot inertia...

138

N. Cuperlier, M. Quoy, and P. Gaussier

Go al m ap

1 MotivA

Pl ac eR ec og ni tio n

0.81

Transition

0.9

MotivB

1

W=0.9

0.81 Ni

Pi

0.9

Fig. 6. Diﬀusion of the activity in the graph corresponding to the cognitive map. Diﬀusion is starting from the goal. Each node keeps the maximal activity coming from its neighbours. The activity is the product between the value of the link and the activity of the sending node.

8 Movement and neural ﬁeld dynamics As seen in section 7, after planning, diﬀerent movements are proposed. a simple competition mechanism selects the neuron with the higher value. But how may these informations be used ? We have ﬁrst tested strict competition between them, but why do not also use other transitions that contains interesting informations about the agent location context ? So we now have several transitions to be taken into account for the movement. The solution used for having a stable continuous direction to follow is to deﬁne a dynamical system where the stable ﬁxed point solution is the direction to follow. This is achieved using a neural ﬁeld [1, 18, 16]. τ.

df (x, t) = −f (x, t) + I(x, t) + h + dt

z∈Vx

w(z).f (x − z, t)dz

(2)

Where f (x, t) is the activity of neuron x, at time t. I(x, t) is the input to the system. h is a negative constant. τ is the relaxation rate of the system. w is the interaction kernel in the neural ﬁeld activation. A diﬀerence of Gaussian (DOG) models these lateral interactions that can be excitatory or inhibitory. Vx is the lateral interaction interval that deﬁnes the neighbourhood. Without inputs the constant h ensures the stability of the neural ﬁeld homogeneous pattern since f (x, t) = h. In the following, the x dimension is an angle (direction to follow according to the north). This equation allows the computation of attractors corresponding to ﬁxed points of the dynamics and to local maxima of the neural ﬁeld activity. A stable direction to follow is reached when the system is on any of the attractors. The angle of a candidate transition is used as input. The intensity of this input depends on the corresponding goal transition activity, but also on its origin place cell recognition activity (see Fig 7). If only one transition is proposed, there will be only one input with an angle xtarg = x∗ and it erects only one

Navigation and Planning in an Unknown Environment

139

Animat Location

B

AB

A

B

BD

A D C

level of place recognition in the neighborhood

AC

global mvt

C

CD

D

level of place recognition in the neighborhood + transitions cells and actions

Fig. 7. The merging mechanism allows to get a better direction (global movement) than the use of the single information obtained from current transition (BD). It takes into account the previous movement performed and the transitions predicted from close enough place cell (C).

attractor x∗ = xtarg on the neural ﬁeld. If xc is the current orientation of the robot, its rotation speed will be proportional to w = x˙ = df (x,t) dt |xc . Merging of several transition informations depends on the distance between them. Indeed if the inputs are spatially close, the dynamics give rise to a single attractor corresponding to the average of them. Otherwise, if we progressively amplify the distance between inputs, a bifurcation point appears for a critical distance, and the previous attractor becomes a repellor and two new attractors emerge. Obstacles are detected by 12 infra-red sensors. A reﬂex behavior is triggered by a Braitenberg-like architecture [6]. When an obstacle is detected on a given direction, the reﬂex system will generate a negative input in this orientation. Hence the bubble of the neural ﬁeld activity will move. Consequently the computed direction will reﬂect this change and allow to avoid this obstacle (see Fig 8). Oscillations between two possible directions are avoided by the hysteresis property of this input competition/cooperation mechanism. It is possible to adjust this distance to a correct value by calibrating the two elements responsible for this eﬀect: spatial ﬁltering is obtained by the convolution of the Dirac like signal coming from motor transition information with a Gaussian and taking it as the input to the system. This combined with the lateral interactions allows the fusion of distinct input as a same attractor. The larger the curve, the larger the merging will be.

9 Conclusion Our model currently running on robots (Koala robots and Labo3 robots) has interesting properties in terms of autonomous behavior. However, this autonomy has some drawbacks:

140

N. Cuperlier, M. Quoy, and P. Gaussier

Robot at ending position

Robot at starting position B

Inputs to the neural field

Location A:

A

Robot trajectory

Obstacle input Neural field

Final movement selection

Neural field

Final movement selection

Joystick direction input

Location B : Inputs to the neural field

Obstacle input

Joystick direction input

Fig. 8. Top: Traectory of a Labo3 robot in an open environment with obstacles. The movement direction is given by a oystick input. We focus on the ﬁnal action selection on two points in this environment: A and B. During all the traectory, the movement ordered by the oystick is go straight. Turn movements in the robot traectory are only due to obstacle avoidance. Middle (A): Neural ﬁeld activity without any obstacle. The direction taken corresponds to the oystick input. Bottom (B): Neural ﬁeld activity with an obstacle. The obstacle shifts the neural ﬁeld maximal activity leading to a turning move.

Navigation and Planning in an Unknown Environment

141

• we are not able to build a cartesian map of the environment because all location learned are robot centered. However, the places in the cognitive map and the directions used give a skeleton of the environment. • we have no information about the exact size of the rooms or corridors. Again, the cognitive map only gives a sketch of the environment. • some parameters have to be set: the recognition threshold (section 3) and the diﬀusion size of the interaction kernel of the neural ﬁeld (section 8). The ﬁrst one determines the density of build places. The higher the threshold, the more places are created. The second parameter has to be tuned for each robot depending on its size and on the position of the infra-red sensors for obstacle avoidance. For instance, a too high diﬀusion value prevents the robot from going through the doors. The transition used in this model may also be the elementary block of a sequence learning process. Thus, we are able to propose a uniﬁed vision of the spatial (navigation) and temporal (memory) functions of the hippocampus [4]. Acknowledgments This work is supported by two french ACI programs. The ﬁrst one on the modeling of the interactions between hippocampus, prefontal cortex and basal ganglia in collaboration with B. Poucet (CRNC, Marseille) JP. Banquet (INSERM U483) and R. Chatila (LAAS, Toulouse). The second one (neurosciences integratives et computationnelles) on the dynamics of biologically plausible neural networks in collaboration with M. Samuelides (SupAero, Toulouse), G. Beslon (INSA, Lyon), and E. Dauce (Perception et mouvement, Marseille).

References 1. S. Amari. Dynamics of pattern formation in lateral-inhibition type neural ﬁelds. Biological Cybernetics, 27:77–87, 1977. 2. M. Arbib and I. Lieblich. Motivational learning of spatial behavior. In J. Metzler, editor, Systems Neuroscience, pages 221–239. Academic Press, 1977. 3. I. Bachelder and A. Waxman. Mobile robot visual mapping and localization: A view-based neurocomputational architecture that emulates hippocampal place learning. Neural Networks, 7(6/7):1083–1099, 1994. 4. J. Banquet, P. Gaussier, M. Quoy, A. Revel, and Y. Burnod. A hierarchy of associations in hippocampo-cortical systems: cognitive maps and navigation strategies. Neural Computation, 17(6), 2005. 5. R. Bellman. On a routing problem. Quarterly of Applied Mathematics, 16:87–90, 1958. 6. V. Braitenberg. Vehicles : Experiments in Synthetic Psychology. Bradford Books, Cambridge, 1984. 7. J. B. C. Giovannangeli, P. Gaussier. Robot as a tool to study the robustness of visual place cells. In I3M’2005: International Conference on Conceptual Modeling and Simulation, Marseille, 2005.

142

N. Cuperlier, M. Quoy, and P. Gaussier

8. J. Donnart and J. Meyer. Learning reactive and planning rules in a motivationnally autonomous animat. IEEE Transactions on Systems, Man and Cybernetics-Part B, 26(3):381–395, 1996. 9. D. Filliat and J.-A. Meyer. Map-based navigation in mobile robots - I. a review of localisation strategies. Journal of Cognitive Systems Research, 4(4):243–282, 2003. 10. P. Gaussier, A. Revel, J. Banquet, and V. Babeau. From view cells and place cells to cognitive map learning: processing stages of the hippocampal system. Biological Cybernetics, 86:15–28, 2002. 11. C. Joulain, P. Gaussier, and A. Revel. Learning to build categories from perception-action associations. In International Conference on Intelligent Robots and Systems - IROS’97, pages 857–864, Grenoble, France, September 1997. IEEE/RSJ. 12. J.-A. Meyer and D. Filliat. Map-based navigation in mobile robots - II. a review of map-learning and path-planing strategies. Journal of Cognitive Systems Research, 4(4):283–317, 2003. 13. J.-A. Meyer and S. Wilson. From animals to animats. In M. Press, editor, First International Conference on Simulation of Adaptive Behavior. Bardford Books, 1991. 14. J. O’Keefe and N. Nadel. The hippocampus as a cognitive map. Clarendon Press, Oxford, 1978. 15. M. Quoy, P. Gaussier, S. Leprˆetre, A. Revel, C. Joulain, and J. Banquet. Lecture Notes in Artiﬁcial Intelligence Series, 1812, chapter A planning map for mobile robots: speed control and paths ﬁnding in a changing environment, pages 103– 119. Springer, ISBN 3-540-41162-3, 2000. 16. M. Quoy, S. Moga, and P. Gaussier. Dynamical neural networks for top-down robot control. IEEE transactions on Man, Systems and Cybernetics, Part A, 33(4):523–532, 2003. 17. A. Revel, P. Gaussier, S. Leprˆetre, and J. Banquet. Planiﬁcation versus sensorymotor conditioning: what are the issues ? In From Animals to Animats : Simulation of Adaptive Behavior SAB’98, pages 129–138, 1998. 18. G. Sch¨ oner, M. Dose, and C. Engels. Dynamics of behavior: theory and applications for autonomous robot architectures. Robotics and Autonomous System, 16(2-4):213–245, December 1995. 19. L. Schwartz. Computational anatomy and functional architecture of striate cortex: a spatial mapping approach to perceptual coding. Vision Res., 20:645– 669, 1980. 20. N. Tinbergen. The study of instinct. Oxford University Press, London, 1951. 21. E. Tolman. Cognitive maps in rats and men. The Psychological Review, 55(4), 1948.

Exploiting Distinguishable Image Features in Robotic Mapping and Localization Patric Jensfelt1 , John Folkesson1 , Danica Kragic1 , and Henrik I. Christensen1 Centre for Autonomous Systems Royal Institute of Technology SE-100 44 Stockholm, SWEDEN [patric,johnf,danik,hic]@nada.kth.se Summary. Simultaneous localization and mapping (SLAM) is an important research area in robotics. Lately, systems that use a single bearing-only sensors have received signiﬁcant attention and the use of visual sensors have been strongly advocated. In this paper, we present a framework for 3D bearing only SLAM using a single camera. We concentrate on image feature selection in order to achieve precise localization and thus good reconstruction in 3D. In addition, we demonstrate how these features can be managed to provide real-time performance and fast matching to detect loop-closing situations. The proposed vision system has been combined with an extended Kalman Filter (EKF) based SLAM method. A number of experiments have been performed in indoor environments which demonstrate the validity and eﬀectiveness of the approach. We also show how the SLAM generated map can be used for robot localization. The use of vision features which are distinguishable allows a straightforward solution to the “kidnapped-robot” scenario.

1 Introduction It is widely recognized that an autonomous robot needs the ability to build maps of the environment using natural landmarks and to use them for localization, [19, 2, 4, 18, 20]. One of the current research topics related to SLAM is the use of vision as the only exteroceptive sensor, [3, 5, 6, 17, 15] due to the low cost. In this paper, we present a SLAM system that builds 3D maps using visual features using a single camera. We describe how we deal with a set of open research issues such as producing stable and well-localized landmarks, robust matching procedure, landmark reconstruction and detection of loop closing. These issues are of extreme importance when working for example in an EKF setting where the computational complexity grows quadratically with the number of features. Robust matching is required for most recursive formulations of SLAM where decisions are ﬁnal.

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 143–157, 2006. © Springer-Verlag Berlin Heidelberg 2006

144

P. Jensfelt et al.

Besides low cost, another aspect of using cameras for SLAM is the much greater richness of the sensor information as compared to that from, for example, a range sensor. Using a camera it is possible to recognize features based on their appearance. This can then simplify one of the most diﬃcult problems in SLAM, namely data association. We demonstrate just how powerful an advantage this is by using the SLAM map to perform robot localization without any initial pose estimate. The main contributions of this work are i) a method for the initialization of visual landmarks for SLAM, ii) a robust and precise feature detector, iii) the management of the measurement to make on-line estimation possible, and iv) the demonstration of how this framework can facilitate loop closing and localization. Due to the low complexity, we believe that our approach scales to larger environments.

2 Related Work Bearing only/single camera SLAM suﬀers from the problem that a single observation of a landmark does not provide an estimate of its full pose. This problem is typically addressed by combining the observations from multiple views as in the structure-from-motion (SFM) approaches in computer vision. The biggest diﬀerence between SLAM and SFM is that SFM considers mostly batch processing while SLAM typically requires on-line real-time performance. The most important problem that has to be addressed in bearing only SLAM is landmark initialization, again because a single observation does not allow all degrees of freedom to be determined. A particle ﬁlter used to represent the unknown initial depth of features has been proposed in [3]. The drawback of the approach is that the initial distribution of particles has to cover all possible depth values for a landmark which makes it diﬃcult to use when the number of detected features is large. A similar approach has been presented in [9] where the initial state is approximated using a Gaussian Sum Filter for which the computational load grows exponentially with number of landmarks. The work in [10] proposes an approximation with additive growth. Similarly to our work, a multiple view approach has been presented in [6]. This work demonstrates the diﬃculties related to landmark reconstruction when the robot performs only translational motion along the optical axis. To cope with the reconstruction problem, a stereo based SLAM method was presented in [17] where Diﬀerence-of-Gaussians (DoG) is used to detect distinctive features which are then matched using SIFT descriptors. One of the important issues mentioned is that their particle ﬁlter based approach is inappropriate for large-scale and textured environments. The contribution of our work is that we deal with this problem using a feature detector that produces fewer features (presented in more detail in Section 4). In our work we use only a few high quality features for performing SLAM.

Exploiting Distinguishable Image Features

145

We have also considered another problem raised in [17] related to time consuming feature matching and use KD-trees to make our matching process very fast. The visual features used in our work are Harris corner features across diﬀerent scales represented by a Laplacian pyramid for feature detection. For feature matching, we take a combination of a modiﬁed SIFT descriptor and KD-trees. In man-made environments, there are many repetitive features and a single SIFT descriptor is not discriminative enough in itself to solve the data association problem. To deal with this problem, the approach using “chunks” of SIFT points to represent landmarks in an outdoor environment has been presented in [12]. This is motivated by the success that SIFT has had in recognition applications where and object/scene is represented as a set of SIFT points. In our approach, the position of a landmark is deﬁned by a series of single modiﬁed SIFT points representing diﬀerent views of the landmark. Each such point is accompanied with a chunk of descriptors that make the matching/recognition of landmarks more robust. Our experimental evaluation shows also that our approach performs successful matching even with a narrow ﬁeld of view which was mentioned as a problem in [6], [17]. One of the more challenging problem in SLAM is loop closing. In [15] visually salient so called “maximally stable extremal regions” or MSERs, encoded using SIFT descriptors, are used to detect when the robot is revisiting an area. A number of approaches have been presented for other sensory modalities, [7]. We also show how our framework can be used for this purpose. In our work, a distinction is made between recognition and location features. A single location feature will be associated with several recognition features. The recognition features’ descriptors then give robustness to the match between the location features in the map and the features in the current image. The key idea is to use a few high quality features to deﬁne the location of landmarks and then use the other features for recognition.

3 Landmark Initialization In this work, we use feature matching across N frames to determine which points make good landmarks. Features that are successfully matched over a number of frames are candidates for landmarks in the map. Having a buﬀer of N frames also allows us to calculate an estimate of the 3D position of the corresponding landmark. The SLAM process is fed with data from the output side of the buﬀer, i.e. with the data from frames that are delayed N steps with respect to the input side of the buﬀer. Fig. 1 illustrates the idea. Since an estimate of the 3D position of landmarks can be supplied with the ﬁrst observation of a landmark presented to the SLAM process (see Fig. 1), the landmarks can immediately be fully initialized in the SLAM process. This allows immediate linearization without the need to apply multiple hypotheses [10] or particle ﬁltering [3] techniques to estimate the depth. It is important

146

P. Jensfelt et al. Matching, Selection & 3D estimation Camera

image(k)

....

meas(k−N) 3D info

SLAM

Fig. 1. Features are tracked over N frames in a buﬀer to allow for quality assessment and feature selection. The 3D position is estimated and landmarks formed from features that are tracked over many frames and are of high quality. The input to the SLAM process is delayed by N frames, but in return an estimate of the 3D position of the points landmarks can be supplied.

to point out that that the approximate 3D position found from the buﬀer of frames is only used to be able to initialize the point landmark at the correct depth with respect to the camera at the ﬁrst observation. In a sense it allows us to know which of the multiple hypotheses about the depth is correct right away which of course saves computations. Having the correct depth allows us, as said before, to handle the linearization errors that would results from having a completely wrong estimate of the depth. In our experience having the SLAM output lagging N frames is not a problem. For cases where the current position of the robot is needed, such as when performing robot control, a prediction can be made using dead-reckoning forward in time from the robot pose given by the SLAM process. For typical values of N , the addition to the robot position error caused by the deadreckoning is small and we believe that the beneﬁts of being able to initialize landmarks using bearing-only information and perform feature quality checks are more signiﬁcant. It is also important to understand that we predict forward the short distance from the SLAM time to the current time in every iteration so the errors from the prediction do not accumulate. When selecting what features to use as landmarks in the map several criteria are considered: i) The feature is detected in more than a predeﬁned number of frames, ii) The image positions of the feature allow good triangulation, and iii) The resulting 3D point is stable over time in the image. The ﬁrst requirement removes the noise and moving targets. The second removes the features that have a high triangulation uncertainty (small baseline, features with bearings near the direction of motion). The third requirement removes features that lack sharp positions in all images due to parallax or a lack of a strong maximum in scale space. Diﬀerence in scales of the images can also cause apparent motion of features, such as for example a corner of a non-textured object. In our implementation the length of the buﬀer, i.e. the number N is ﬁxed. We have tested with values between 10 and 50. Since one of the key purposes with the buﬀer is to allow for 3D estimation we demand that the camera has moved to add a new frame to the buﬀer. This way, it is likely that there is a baseline for estimating the location. The value of N depends very much on the motion of the robot/camera and the camera parameters. For a narrow ﬁeld

Exploiting Distinguishable Image Features

147

of view camera mounted in the direction of motion of the robot as in our case the eﬀective baseline will be quite small. As part of our future work we plan to compare the results that we get from using an omnidirectional camera.

4 Feature Description It has been shown in [14] that the SIFT descriptor [11] is the most robust regarding scale and illumination changes. The original version of the SIFT descriptor uses feature points determined by the peaks of a series of Diﬀerence of Gaussians (DoG) on varying scales. In our system, peaks are found using Harris-Laplace features, [13] since they respond to regions of high curvature, instead of blob-like image structures obtained by series of DoG. This leads to features accurately localized spatially, which is essential when features are used for reconstruction and localization, instead of just recognition. In addition, Harris-Laplace generates on average one fourth of points[8] compared to the regular DoG approach which is an important beneﬁt for SLAM where we strive to keep the number of features low for computational reasons. The original SIFT descriptor assigns canonical orientation at the peak of smoothed gradient histograms. This means that similar corners but with a signiﬁcant rotation diﬀerence can have similar descriptors. In a sparse, indoor environment many of the detected features come from corner features and there may potentially be many false matches. For example, the four corners of the waste bin in Fig. 2 might all match if rotated. Therefore, we use a rotationally ’variant’ SIFT descriptor where we avoid the canonical orientation at the peak of smoothed gradient histogram and leave the gradient histogram as it is.

Fig. 2. Example image from an indoor environment showing that many corner structures are very similar if rotated. Examples are the corners of the waste bin and the window sections on the door.

148

P. Jensfelt et al.

5 Feature Tracking As outlined in Section 3 a buﬀer with the last N frames is stored in memory. The image data itself does not need to be stored, it is the feature points extracted from the frames that are stored. The feature points are tracked in consecutive image frames as the robot moves. The similarity between two feature points is the distance between the descriptors which are vectors in a 128-dimensional space. To manage the matching between frames, lists with association/points are maintained. Fig. 3 shows the basic organization of this frame memory. On the left we have the buﬀer with the N frames drawn vertically with the input side at the top and output side that is fed into the SLAM process at the bottom. On the right side of the buﬀer in the frame memory we show the association list that keep track of which feature points in the diﬀerent frames have been matched. Each such association list item corresponds to one landmark in the world. The SIFT descriptor of a feature changes when the camera moves. However given that the feature is continuously tracked this change in descriptor is typically small between consecutive frames and can be handled. Each association/point in the world, pi , will have a set of descriptors dj corresponding to diﬀerent vantage points for each of the tracked features over the N frames. These lists can be analyzed to judge the quality of the corresponding landmark candidate as described in Section 3. The output from the tracking module is a selection of the points in the oldest frame in the buﬀer. These points correspond to either already initialized landmarks or new landmarks of high quality that can be initialized. Along with new landmarks an estimate of the 3D position is provided. This estimate is only used in the initialization step and thus only have to be performed for new landmarks. The number of points that are output is only a small fraction of all points detected in a frame. To make the matching procedure faster and more robust, we predict the approximate image location of the features from previous frames using odometry and optical ﬂow estimates. The buﬀer allows us to predict the position of features detected not only in the previous frame but also in frames before that. We make one prediction for each potential landmark in the old frames, i.e., we only use the last of features that have been matched between frames. One important observation in our investigation was that even with very small changes between frames the same features would typically not be detected in every frame which would mean loosing track of most features if only matching against the previous frame. This observation is true also when using the DoG detector for the features. Using the prediction allows us to reduce the search window when looking for matches between frames. Features that do not match the predicted positions of points currently in the frame memory are matched to landmarks in the database. This also allows for fast detection of potential loop closing situations. The ﬁrst time the location of a landmark has been established through triangulation and it has met the other criteria listed in Section 3, it is added

Exploiting Distinguishable Image Features

149

to the database as a new location feature. This is discussed further in the next section. DATABASE

... ... ...

... Output frame

Landmarks

FN

... ...

.......

.......

...

F2

... ...

...

N

F1

...

.......

Recognition

... ...

Input frame

Location

Associations/ Points

...

FRAMEMEMORY Frames

Fig. 3. A schematic view of the frame memory that stores the points and the associations between points in the N last frames and the database. The point in the frame memory that have been matched or added to the database have a reference to the corresponding landmark there. Each landmark in the database has a set of descriptors that corresponds to location features seen from diﬀerent vantage points. To validate a match, each of these descriptors keeps a list of the other descriptors found in the same frame. We refer to these as recognition descriptors. These provide the ability to “recognize” a landmark again.

6 Database Management As the robot moves along in the environment, features will leave its ﬁeld of view and thus eventually also the frame memory described in Section 5. In SLAM, it is important for the robot to detect if it is revisiting an area and ﬁnd the correspondence between the new features in the image and existing landmarks. In fact, this is an issue not only when revisiting an area after closing a large loop but also when turning the camera back to an area that has not been looked at for a while. To handle this we use a database that stores information about the appearance of the landmarks from diﬀerent view points. That is, we let the SLAM process take care of estimating the position of the landmarks and leave it to the data base to deal with the appearance and matching of them.

150

P. Jensfelt et al.

In our database, each landmark has a set of SIFT descriptors that correspond to diﬀerent vantage points. These descriptors are provided by the frame memory that matches the image points between frames and stores this in the association/point lists. A new descriptor is added to a landmark when it differs more than a certain threshold from all other descriptors attached to that landmark. Fig. 3 shows the structure of the database where the landmarks are denoted with F1 , F2 , . . . , FN . The dashed box contains the descriptors for each of the landmarks. A KD-tree representation and a Best-Bin-First [1] search allow for real-time matching between new image feature descriptors and those in the database. There may be potentially many similar descriptors corresponding to diﬀerent features. Looking once again at the image in Fig. 2 the four corresponding corner points on the glass windows on the left hand side door section will all be very similar. Trying to ﬁnd a correspondence between images features in a new frame and the map landmarks with a bit of uncertainty induced by a loop can be hard based only on a single SIFT descriptor. Therefore, we require multiple matches to increase the robustness [6]. We refer to these multiple descriptors as recognition descriptors and the corresponding image features as recognition features. That is, when matching a feature to the database we ﬁrst look for matches between its location descriptor1 and the descriptors in the database. Then, we verify the match using the corresponding two sets of recognition descriptors. As an additional test, it is required that the displacement in image coordinates for the descriptor is consistent with the transformation between the two frames estimated from the matched recognition descriptors. Currently, the calculation is simpliﬁed by checking the 2D image point displacement. This ﬁnal conﬁrmation eliminates matches that are close in the environment and thus share recognition descriptors such as would be the case with the glass windows in Fig. 2. To summarize, the matching of new features with the database proceeds as follows: 1. Find matching candidates by matching with the set of location descriptors in the database (dashed box in the database in Fig. 3). The KD-tree representation allows for fast matching and an eﬀective way to eliminate all but a few of the potential location feature matches in the database. 2. Validate the candidates using all extracted descriptors from the current frame, i.e. the recognition feature and the recognition features associated with the matches from step 1. 3. Conﬁrm candidate by checking that the motion given all the matches is consistent. 1

Each feature in a new image is a potential location feature and the rest of the features in that frame will be its recognition features. A feature can thus be both a location feature and act as recognition feature for one or more other location features in that frame.

Exploiting Distinguishable Image Features

151

7 Using the Location Features for SLAM and Localization We have seen in the previous sections how features are tracked between frames using a buﬀer of N frames and how keeping this buﬀer allows for judging the quality of potential landmarks and for ﬁnding an estimate of the 3D position of the landmarks before they are initialized in SLAM. We use an EKF based implementation for SLAM but the output from the frame memory and database can be fed into any number of SLAM algorithms. The main diﬀerence between the EKF implementation used here for SLAM and the standard formulation is that we supplement the ﬁrst bearing-only measurement of a new landmark with the additional information about the approximate distance as determined through the triangulation in the frame buﬀer. The distance information s not given in the form of an EKF measurement, these are always only bearings. It is used more like some oracle told the SLAM ﬁlter where in depth to initialize a new landmark. As a result an accurate linearization can be made in the EKF as of the ﬁrst sighting without the need for any special arrangement to account for an unknown depth as in e.g. [3, 10]. Once the map is built, it can be used for localization. Since the landmarks in the database are distinguishable, it allows the robot to recognize areas that are part of its map using a single landmark. Thus the robot can automatically initialize itself in the map after which the map can be used to track the robot pose. This is similar to the situation when closing a loop and having to make the association between new measurements and old landmarks in the map. In the latter case one has a rough idea about where to look for such a match. When performing global localization the uncertainty is the entire space and the search for matches can thus become quite expensive. The framework presented in this work allows for fast matching against the database which we will further investigate in the experimental section. One of the beneﬁts of our representation with multiple descriptors for the landmarks, each of them encoding the appearance from a certain vantage point, is that this information can be used to deduce the approximate position of the robot from a single point observation. To allow for this, the pose from which each descriptor was ﬁrst observed is stored along with the descriptor. The idea is not that to get an exact position out of an observation but to narrow down the area that the robot is likely to be in. To ﬁnd the pose of the robot in the map we initialize a particle ﬁlter as soon as a match to the database is found. The particles are initially distributed around the pose indicated by the database match. The orientation for each particle in the initial sample set is given by the bearing angle of the observation. This cuts the unconstrained degrees of freedom down to two. Given that the 3D position of the landmark is given, that the ﬂoor is ﬂat and that a measurement of the bearing to the landmark is given not only in the xyplane but also in the vertical direction the distance to the landmark can also be estimated from a single measurement. In our current implementation have

152

P. Jensfelt et al.

not incorporated this last bit of information. After the ﬁrst observation it will then require only a few other database matches to reduce the spread of particles enough to say that the position is known. In our current implementation we initialize a normal EKF localization algorithm at this point.

8 Experimental Evaluation The experimental evaluation has been carried out on a PowerBot platform from ActivMedia. It has a diﬀerential drive base with two rear caster wheels. The camera used in the experiments is a Canon VC-C4 CCD camera. This camera has built-in pan-tilt-zoom capabilities but in the experiments these were all ﬁxed with a slight tilt upwards to reduce the amount of ﬂoor in the image. The non-holonomic motion constraints of the base makes it hard to generate large baselines for triangulation as the motion mostly is along the optical axis. This in combination with the relatively narrow ﬁeld of view2 contributes to the diﬃculty of the problem.

Fig. 4. The experimental PowerBot platform with the Canon VC-C4 camera mounted on the laser scanner.

The experimental evaluation is carried out in two steps. In the ﬁrst step we let the robot build up a database of visual landmarks while feeding them into SLAM to build a map with the 3D position of these landmarks. This test will show that i) our framework produces few but stable landmarks well suited for map building ii) we can match new observation to the database when closing the loop. 2

The ﬁeld of view of the Canon VC-C4 is about 45◦ in the horizontal plane and 35◦ vertically when the camera is in “wide-angle mode” as in our experiments.

Exploiting Distinguishable Image Features

153

In the second step the robot is given the task to ﬁnd its position in this map given the database and the map. This test will underscore the ability to match observations to the database without false matches and highlight the strength our representation provides to a localization system. The setting for the experiments is an area around an atrium that consists of loops of varying sizes. In the map building experiment we let the robot drive 3 laps around a part of the environment, each lap being approximately 30m long. This trajectory is shown along with a map built using a laser scanner in Fig. 5 (dark “circular” path). The experiment took 8 minutes and 40 seconds and the time to process the data was 7 min 7s on a 1.8GHz laptop computer which shows that the system can run in real-time even if all unmatched features are matched to the database in every processed frame. After the ﬁrst loop the database/map consisted of 98 landmarks. The landmarks are shown as red/dark squares in Fig. 5. After the 3 loops were completed the map consisted of 113 landmarks. This shows that the robot was successfully matching most of the observations to the database during the last two laps. Otherwise one would expect the map to be roughly 3 times the size after the ﬁrst loop. There are two important characteristics to note about the map with 3D visual landmarks. Firstly, there are far fewer features then is typically seen in other works where SIFT like features are used in the map [17, 16]. This can be attributed to using only the most stable points features as SLAM landmarks and the rest for recognition/matching of those landmarks. Realtime performance was not demonstrated in [17, 16]. Secondly, the landmarks are well localized which can be seen by comparing with the walls from the superimposed map built using the laser scanner3 . Some of the landmarks are lamps hanging from the ceiling such as the one at the upper right corner of the robot path in Fig. 5. This lamp is also visible in the upper image on the right side of the same ﬁgure. The area to the right in the map with a large number of landmarks is shown in the lower right image in Fig. 5. This area has many objects at diﬀering depths. In the back against the walls there are ﬁve heavily textured paintings. The line in the laser based map comes from the benches that are in front of the wall which accounts for the line of visual landmarks behind the line made using the laser. Part of the spread is also due to uncertainty in depths of the landmarks. The robot moved close to orthogonal to the wall creating very little baseline in the data fed into SLAM. It is worth while repeating that the depth estimate from the frame memory only serves to get the initialization of the depth roughly right to reduce the linearization errors but that the real estimate of the 3D position is calculate through the SLAM process using the bearing-only measurements. Depending on the scene, a typical frame has between 40-100 point features. The time to perform the tracking over frames has constant complexity. Out of 3

Note that the laser based map is only shown for reference. Only vision was used in the experiments

154

P. Jensfelt et al.

Fig. 5. The SLAM map is shown as red squares at the locations of landmarks. To help visualize the environment, a separately made laser scanner map is superimposed on this map. Also shown are the trajectories from when the map was built (3 loops) and when the robot was trying to localize. The feature match between the evening database image image (top) and the daytime localization experiment (middle) is indicated. The bottom image corresponds to the area of the map where the density of landmarks is greatest.

the features in each new frame as many as half typically do not match any of the old features in the frame memory and are thus matched to the database. The matching to the database uses the KD-tree in the ﬁrst step which makes this ﬁrst step fast. This often results only in a few possible matching candidates. A typical landmark in the database has around 10 descriptors acquired from diﬀerent viewing angles. As we noted while building this map, the landmarks inserted into the database during the ﬁrst loop were then matched and updated on the second two loops with no diﬃculty. In order to demonstrate the usefulness of this result we then used the map and database in the second experiment to localize the robot at a diﬀerent time of day, when there was sunlight streaming through the windows and students milling about. When performing localization we do not need to estimate the 3D location of the points and thus the output from the frame memory is shifted to be at the input side, i.e., without the delay. The

Exploiting Distinguishable Image Features

155

frame memory is still used to track the points over time but not to verify the quality which is most important when building the map. The robot was given no initial pose information relative to the database and map. In Fig. 5 the path of the robot in this second experiment is overlayed. The robot initially moves around in the lower left corner of the ﬁgure before it moves to the right for some time and then eventually moves up into regions that were mapped in the ﬁrst experiment. The robot moves in unmapped areas for almost 6 minutes. During this time it is constantly trying to ﬁnd matches between unmatched features in the image to the database. More than 1,000 images are added to the frame memory and matched without any false matches to the database. When the robot enters an area were it can observe landmarks from the database/map it almost immediately recognized a landmark and a particle ﬁlter consistent with the observation was initialized. In seconds, after repeated observations of several landmarks, the particles had converged enough to initialize the EKF localization. The evolution of the particle ﬁlter from initialization to convergence is shown in Fig. 6 in four frames from left to right. In the last frame in this ﬁgure, the precision of the localization is indicated by the alignment of the laser scan (not used for localization, only for comparison) and the walls in the map.

Fig. 6. Here we show the evolution of the localization particle ﬁlter. On the left we see the initial distribution after ﬁrst match to the database. On the right is shown the robot after initialization in the map. The laser scan is also shown in the right image to help conﬁrm that the localization is indeed correct.

The problems of a narrow ﬁeld of view for an ordinary camera were overcome in these experiments partly by carefully driving the robot. Better results were observed when the robot was driven so that features remained in view for as long as possible, (notice the rather crooked paths in Fig. 5).

156

P. Jensfelt et al.

9 Conclusion and Future Work The contributions of the framework presented here are the feature selection and matching that allows for real-time vision based bearing-only SLAM. We distinguish between location and recognition features. The location features correspond to points in 3D for which the robot motion allows good triangulation and which are used as landmarks in the map. The matching is made robust by the inclusion of many recognition features from the image for each location feature. The use of Harris-Laplace corner detection combined with a scale space maximization gives rotationally variant features which are more appropriate for the camera motions generated by a camera mounted on a mobile robot. We use three criteria to select features for SLAM: persistence, range estimate quality and image stability. The persistence criteria eliminates spurious and dynamic features by requiring that the feature be observed many times. The quality of the range estimate depends on the motion of the robot relative to the 3D point. If the motion does not produce a suﬃcient baseline there is no reason to use the associated vision feature. The stability in the image depends on how well the feature is localized in the image. The evaluation of our vision based landmarks was done on data collected from a robot moving through a indoor environments. We were able to report no false matches and the creation of accurate 3D maps. We also showed that those maps could be used to localize the robot automatically in the environment. One topic for future research is to actively control the pan-tilt degrees of freedom of the camera on the robot. This would allow the robot to focus on a landmark for a longer time and create a better baseline and thus a more accurate map. Another way to address the problem with the limited ﬁeld of view is to use an omnidirectional camera which is also part of the planned future work.

References 1. J. S. Beis and D. G. Lowe. Shape indexing using approximate nearest-neighbour search in high-dimensional spaces. In IEEE CVPR, pages 1000–1006, 1997. 2. J. A. Castellanos and J. D. Tard´ os. Mobile Robot Localization and Map Building: A Multisensor Fusion Approach. Kluwer Academic Publishers, 1999. 3. A. J. Davison. Real-time simultaneous localisation and mapping with a single camera. In ICCV, 2003. 4. G. Dissanayake, P. Newman, S. Clark, H.F. Durrant-Whyte, and M. Corba. A solution to the slam building problem. IEEE TRA, 17(3):229–241, 2001. 5. J. Folkesson, Jensfelt P, and H. I. Christensen. Vision slam in the measurement subspace. In IEEE ICRA05, 2005. 6. L. Goncavles, E. di Bernardo, D. Benson, M. Svedman, J. Ostrovski, N. Karlsson, and P. Pirjanian. A visual fron-end for simultaneous localization and mapping. In IEEE ICRA, pages 44–49, 2005.

Exploiting Distinguishable Image Features

157

7. J. Gutmann and K. Konolige. Incremental mapping of large cyclic environments. In IEEE Int. Symposium on Computational Intelligence in Robotics and Automation, volume 1, pages 318–325, 1999. 8. Patric Jensfelt, Danica Kragic, John Folkesson, and M˚ arten Bj¨ orkman. A framework for vision based bearing only 3D SLAM. In Proc. of the IEEE International Conference on Robotics and Automation (ICRA’06), Orlando, FL, May 2006. 9. N. M. Kwok, G. Dissanayake, and Q.P. Ha. Bearing only SLAM using a SPRT based gaussian sum ﬁlter. In IEEE ICRA05, 2005. 10. T. Lemaire, S. Lacroix, and J. Sol` a. A practical 3D bearing-only SLAM algorithm. In IEEE/RSJ IROS, pages 2757–2762, 2005. 11. David G. Lowe. Object recognition from local scale-invariant features. In ICCV, pages 1150–57, 1999. 12. R.H. Luke, J. M. Keller, M. Skubic, and S. Senger. Acquiring and maintaining abstract landmark chunks for cognitive robot navigation. In IEEE/RSJ IROS, 2005. 13. K. Mikolajczyk and C. Schmid. Indexing based on scale invariant interest points. In IEEE ICCV, pages 525–531, 2001. 14. K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. In IEEE CVPR, pages 257–263, 2003. 15. P. Newman and K. Ho. SLAM-loop closing with visually salient features. In IEEE ICRA, pages 644–651, 2005. 16. S. Se, D. G. Lowe, and J. Little. Mobile robot localization and mapping with uncertainty using scale-invariant visual landmarks. IJRR, 21(8):735–58, 2002. 17. R. Sim, P. Elinas, M. Griﬃn, and J. J. Little. Vision-based slam using the raoblackwellised particle ﬁlter. In IJCAI Workshop on Reasoning with Uncertainty in Robotics, July 2005. 18. J.D. Tard´ os, J. Neira, P.M. Newman, and J.J. Leonard. Robust mapping and localization in indoor environments using sonar data. IJRR, 4, 2002. 19. S. Thrun, D. Fox., and W. Burgard. A probalistic approach to concurrent mapping and localization for mobile robots. Autonomous Robots, 5:253–271, 1998. 20. S. Thrun, Y. Liu, D.Koller, A. Ng, Z. Ghahramani, and H. Durrant-White. SLAM with sparse extended information ﬁlters. IJRR, 23(8):690–717, 2004.

Image-Based Visual Servoing for Mobile Robots with Catadioptric Camera Gian Luca Mariottini, Domenico Prattichizzo, and Andrea Cerbella Dipartimento di Ingegneria dell’Informazione, Universit` a di Siena, Via Roma 56, 53100 Siena, Italy. {gmariottini,prattichizzo}@ing.unisi.it Summary. This paper presents an image-based visual servoing strategy for holonomic mobile robots equipped with a catadioptric camera. This kind of vision sensor combines lens and mirrors to enlarge the ﬁeld of view. The proposed visual servoing is based on the auto-epipolar property, a special conﬁguration which occurs when the desired and the current views undergo a pure translation. This occurrence is detectable from a set of image points, observing when the so-called bi-conics have a common intersection. The proposed visual servoing is divided in two sequential steps. Lyapunov-based stability demonstrates the parametric robustness of the proposed method. Experimental results are presented to show the applicability of our strategy in a real context.

1 Introduction In this paper we propose an image-based visual servoing (IBVS) strategy for the autonomous navigation of a holonomic mobile robot. Both the current and the desired robot conﬁgurations are only speciﬁed through a set of features points acquired by the on-board omnidirectional camera. Diﬀerently to position-based visual servoing (PBVS), in IBVS both the control objective and the control laws are directly expressed in the image domain. As a consequence, such schemes do not need any a priori knowledge of the 3D structure of the observed scene points. Moreover designing the feedback at the sensor level improves the system performances especially when uncertainties and disturbances aﬀect the model of the robot and the camera calibration [8]. However, system convergence can typically be guaranteed only in a neighborhood of the desired conﬁguration. In [10] a hybrid (2-1/2 D) visual servoing approach for articulated manipulators has been presented. It combines IBVS and PBVS features and guarantees convergence in the whole task space. In particular camera rotation and translation between the current and the desired views are directly estimated from the observation of planar objects and then used in the control law.

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 159–170, 2006. © Springer-Verlag Berlin Heidelberg 2006

160

G.L. Mariottini, D. Prattichizzo, and A. Carbella

While most of these works deal with articulated manipulators, there has been an increasing interest in the PBVS control of mobile robots, as in [16, 7, 2]. One of the main problems in visual servoing is that of keeping the features in the camera ﬁeld of view (FOV). This problem is naturally solved if a catadioptric camera is used in the servo loop since it provides panoramic FOV. Catadioptric cameras consist of a coupling between mirrors (catoptric elements) and conventional cameras (dioptric elements) [13]. Some PBVS strategies using catadioptric cameras have been presented [4, 3, 17]. In recent years there has been a great interest in the vision-based control with central catadioptric cameras. However, some of these works assume known the camera-feature distance (in meters)[12, 1]. The original contribution of this work is the design of a IBVS algorithm for holonomic mobile robots with a catadioptric camera on it, that does not need for any 3-D scene metrical information and that does need of any multiple-view estimation process. Our control algorithm is based on the execution of two sequential steps. The ﬁrst one compensates the orientation error, i.e., aligns the robot to the desired conﬁguration by exploiting the auto-epipolar property for panoramic cameras (see Sect. 2). The second step zeroes the translational displacement using feature points to drive the robot to the target. The paper is organized as follows. In Section 2, the auto-epipolar property is presented. Section 3 introduces the image-based control law. Section 4 describes some experimental results. Conclusions are presented in Section 5.

2 Auto-epipolar Conﬁgurations This section presents the auto-epipolar property for panoramic cameras. The auto-epipolar property [6] has been ﬁrst applied by [15] for vision-based control with pinhole cameras. Consider the case in Fig. 1(a) with a catadioptric camera made of a parabolic mirror centered at f and an orthographic camera. By construction, every scene point p ∈ IR3 is projected onto the mirror surface at x ∈ IR3 through the focus f . The image point m (pixels) is then obtained via orthographic projection of x. For a more detailed description of the imaging model for catadioptric cameras the reader can refer to [5]. Consider now two panoramic views as in Fig. 1(b), acquired in fc and fd by the same camera and referred to as {c} current and {d} desired, respectively. Without loss of generality, we choose the world reference frame centered at the desired camera frame in fd . This allows us to express all image entities directly in the mirror reference frame onto the mirror surface, if not otherwise speciﬁed. Given a pair of views of a scene point pi , there exists a matrix E ∈ IR3×3 , called essential matrix [6], such that xdi T Exci = 0

where

E = tR

i = 1, ..., N

(1)

IBVS for Mobile Robots with Catadioptric Camera p

image plane

p1

p2 π2

O ym

X xm zm

P

parabolic mirror

161

cc2

xc2

xc1 fc

Cc

(a)

π1 xd2 xd1 cd1

cc1

fd

R,t

cd2

Cd

(b)

Fig. 1. (a) Imaging model of central catadioptric camera. (b) Basic epipolar geometry setup for central catadioptric cameras.

for all N corresponding mirror points xci and xdi obtained via back-projection of the corresponding image points mdi and mci . t is the skew-symmetric matrix obtained from t. It is an easy matter to show that when R = I, i.e., pure translations, then the essential matrix is skew symmetric, i.e., E = t. Other conﬁgurations exist (such as rotations of π around t) providing a skew symmetric essential matrix, as pointed out in [9], but they will not occur in our planar motion case. Deﬁnition 1 (Auto-epipolar conﬁgurations). The relative conﬁgurations between two cameras yielding to a skew-symmetric essential matrix are referred to as auto-epipolar conﬁgurations. Remark 1. We now deﬁne the new concept of bi-conic, that will allow us to detect the auto-epipolar conﬁgurations for catadioptric cameras directly in the image plane and without any estimation. Deﬁnition 2 (Bi-conics). Consider two projections of the same scene point p onto two corresponding image points md and mc . Overlap the two images. Let xd and xc be their back-projections to their respective mirror surfaces. The bi-conic is the intersection between the plane ψ (with normal vector n = xd × xc ) and the desired camera mirror1 (see Fig. 2). Proposition 1 (Common intersection line). Consider a set of N corresponding points mdi and mci , i = 1, ..., N and let xdi and xci be their backprojections onto the mirror. Consider the matrix M ∈ IRN ×3 containing all the normal vectors ni = xdi × xci of the planes ψi , namely M = [n1 , ..., nN ]T . If the matrix M is rank 2, then all planes ψi intersect at a common line ∈ IR3 or, equivalently, xdi T ˆxci = 0, ∀ i. In other terms, 1

is perpendicular to ni , ∀i.

Note that xc and xd belong to the bi-conic.

162

G.L. Mariottini, D. Prattichizzo, and A. Carbella p

image plane

md mc

xd catadioptric mirror

fd

xc fc

xc

xd

n

ψ bi-conic

Fig. 2. Construction of a bi-conic through the superimposition of xc and xd . xc and xd deﬁne a plane with normal n that intersects the mirror at the bi-conic.

Proof. If M is rank 2, then there exists a right null vector such that M = 0, namely (xdi ×xci )T = 0 ∀i. This means that each plane ψi contains the vector and since all planes pass through the origin of the desired mirror frame, then there exists a unique common intersection line between all planes. The last equivalence comes directly from the triple scalar product property: the expression (xdi × xci )T = 0 is equivalent to xdi T ˆxci = 0 ∀i. In conclusion, if N diﬀerent planes ψi all share a common intersection line, then the bi-conics intersect at two common points. The relationship between bi-conics and auto-epipolar conﬁgurations is stated in the following theorem. Theorem 1. Consider N corresponding point pairs (mdi , mci ) in two views and let xdi and xci be their back-projections onto the mirror surfaces. Assume T that the matrix A = [a1 , a2 , . . . , aN ] is of rank 8, where ai = (xci ⊗ xdi )T , being the symbol ⊗ the Kronecker product and the resulting vector ai ∈ IR9×1 . The N planes ψi with normal ni = xdi × xci all intersect at only one line if and only if the essential matrix E is skew symmetric. Proof. (⇒) If all planes ψi with normal vector ni = xdi × xci intersect at a T single line then the matrix M = [n1 , ..., nN ] is rank 2. This is equivalent to say that there exist a subspace S of dimension equal to one (i.e. a line) that belongs to all planes. In other terms ∃ ∈ S : (ni )

T

= 0 ∀ i = 1, 2, ..., N.

IBVS for Mobile Robots with Catadioptric Camera

163

w

cc2

y(0)

cc1

fc

θ

cc2

y(0)

cc1

fc

cc2

y(t)

t

cc1

fc

v=[vx,vy] cd1 cd2

fd

cd1

cd1 x(0)

Starting configuration

cd2

fd

x(0)

cd2

fd

x(t)

Second Step Translational motion

First Step Auto-epipolar configuration Desired features

No intersection

s(t)

Current features

Fig. 3. The visual servoing algorithm steers the holonomic camera-robot from the starting conﬁguration centered at fc (no common intersection between bi-conics exists), through an intermediate conﬁguration when both views have the same orientation (common intersection exists), towards the position fd (s(t) = 0).

From the triple scalar product property, this is equivalent to xdi T ˆxci = 0 ∀i. Being E unique for the full rank hypothesis on A [9], then E = ˆ, i.e., E is skew symmetric. (⇐) Being rank(E) = 2, then ∃ w = 0 s.t. Ew = 0. If E is skew symmetric, then we distinguish two cases [9]: E1 = t and E2 = teuˆπ (where u = t/||t||). If we consider E1 then the epipolar constraint reads as: xdi T E1 xci = (xdi × xci )T t = 0

∀ i = 1, ..., N.

(2)

It is an easy matter to show that E2 = −t and then the same epipolar constraint of (2) is veriﬁed also for E2 . Equation (2) is equivalent to impose that all the planes ψi with normal vector ni intersect at a common line t, i.e., the baseline. Fig. 3 (bottom-left) shows the image plane bi-conics when the planes ψi do not share a common intersection line, i.e., the cameras are not in autoepipolar conﬁguration (Fig. 3 (upper-left)). On the other hand Fig. 3 (bottomcenter ) shows the image plane bi-conics when all planes ψi have a common intersection line and the cameras are in auto-epipolar conﬁguration (Fig. 3 (up-center )). The previous theorem and proposition provide an important tool for recognize when the cameras are in auto-epipolar conﬁguration. It is worthwhile to mention that to build matrix M we only need to know backprojected mirror points xci and xdi that can be retrieved from the image points

164

G.L. Mariottini, D. Prattichizzo, and A. Carbella 0.04 0.035 0.03

Value ξ

0.025 0.02 0.015 0.01 0.005

-150

-100

-50

0 50 Angle θ [deg.]

100

150

2

Fig. 4. Singular value ratio ξ = σσ32 with respect to the current camera orientation variation. ξ resulted to be convex in many simulation results.

mai and mdi without any time consuming estimation process. It is paramount to note that in the case of camera planar motion, the auto-epipolar condition can be checked when at least N = 3 corresponding points are observed [9], also in the case of coplanar 3D scene points. To test if M is rank deﬁcient we exploit the smallest singular value σ3 obtained by means of SVD of M . As the displacement approaches to a pure translation, then the smallest singular value approaches to zero. In practical cases, however, we experimentally veriﬁed that better results are obtained if the smallest singular value is squared and normalized with respect to the second singular value σ3 2 ξ= . (3) σ2 If the two cameras are purely translated then ξ is exactly zero, as shown in Fig. 4. Remark 2. Function ξ will be used as a control variable for the rotational part of the proposed visual servoing. Note that the derivative of ξ is an odd function in a very large neighborhood of θ = 0◦ (Fig. 4).

3 The Image-Based Control Law In this section we present an image-based visual servoing control law for holonomic mobile robots equipped with a catadioptric camera. This control law (1st step) is based on the auto-epipolar property for panoramic cameras presented in the previous section. Let q(t) = [x(t) y(t) θ(t)]T be the robot conﬁguration vector with respect to the world frame Oxyz w . The holonomic robot kinematic model can be written as

IBVS for Mobile Robots with Catadioptric Camera

x˙ = vx y˙ = vy θ˙ = ω

165

(4)

where v = [vx vy ]T is the translational velocity and ω is the rotational velocity. Suppose that a catadioptric camera is ﬁxed on the mobile robot such that its optical axis is perpendicular to the xy-plane. We suppose that a target image, referred to as desired, has been previously acquired in the desired conﬁguration qd = [0 0 0]T . The current image is available at each time instant from the camera-robot in the current position. As shown in the initial setup of Fig. 3 (upper-left), the robot disparity between the current and the desired poses is characterized by a rotation of an angle θ and a translation t ∈ IR2 . The control law will be able to drive to zero this disparity only using visual information (i.e., θ → 0 and ||t|| → 0). In particular, we will regulate separately the rotational disparity (ﬁrst step) and the translational displacement (second step). 3.1 First Step: Rotation Compensation The goal of the ﬁrst step is to drive the current camera-robot to have the same orientation of the desired one. This condition is simply detectable when ξ = 0 in (3). During this step, the translational velocities are set to zero vx = vy = 0, while the angular velocity is set to ω = −αω σs (t)ξ

(5)

˙ and ξ is the control variable where αω is a positive scalar, σs (t) = sign(ξ) introduced in (3). The local asymptotic convergence of this control law has been proven in [11] by means of the Lyapunov function V = 12 θ2 . Although the convergence is local, the domain of attraction for θ is wide. Many simulative and experimental results shown that convergence is guaranteed with respect to θ in a very large set of angles ([−150◦ , 150◦ ]) around the equilibrium θ = 0◦ . 3.2 Second Step: Translation Compensation Let us now address the second step that will start when ||ξ(t)|| = 0, i.e., when only a pure translation is needed to reach the target conﬁguration. Actually, in the experiments, a user-deﬁned threshold ε is considered. During this translational motion to the desired position, all bi-conics Cdi intersect and the corresponding feature points mci will slide onto the conic Cdi toward mdi . The translation will stop when all features match, i.e., when s(t) = si (t) is equal to zero, being si (t) the arc-length of the conic Cdi from mci (t) to mdi (t) (see Fig. 3 bottom-right). The direction of translation to the desired conﬁguration fd can be retrieved from the knowledge of the epipole cd pointing

166

G.L. Mariottini, D. Prattichizzo, and A. Carbella

toward fc , chosen as the intersection between bi-conics. In [11] it has been shown that v = −s(t)[ cdx cdy ]T . (6) guarantees the robot converges to the target conﬁguration.

4 Experiments In order to validate the proposed image-based visual servoing, a series of experiments have been carried out at the SIena Robotics and Systems lab (SIRS lab). The experimental setup consists of a NOMAD XR4000 holonomic robot with a catadioptric camera mounted on. The catadioptric camera is made of a CCD color camera LU175C by Lumenera screwed on a folded catadioptric mirror NetVision360 by Remote Reality. The whole catadioptric camera can be modeled by the equation x2 +y 2 zm + a2 = m2a m (Fig. 1(a)) where a = 33.4 mm. The orthographic camera parameters are fx = 13 mm, fy = 14 mm, u0 = 616.3 and v0 = 628.2 pixels. All image acquisition and processing algorithms are realized using Intel’s OpenCV libraries on a notebook endowed with a 2 GHz Pentium 4 processor. A set of n = 3 feature points have been selected in the image in order to show the applicability of our technique with the theoretically minimum number of points. This implies that we have three bi-conics on the image plane as shown in Fig. 5. The initial conﬁguration of the camera-robot is set to qini = [1 m 1 m π/4 rad] while the desired one is set to qdes = [0 0 0] as shown in Fig. 5(a) where the desired and the initial robot conﬁgurations are superimposed. In the initial conﬁguration being the orientation not equal to zero bi-conics do not intersect at the same two points (Fig. 5(b)). First, the rotational controller in Sect. 3.1, based on the auto-epipolar property, is used to compensate the rotation (see Fig. 5(c)) with respect to the desired conﬁguration, until all bi-conics intersect, as in Fig. 5(d). Note that in real experiment, notwithstanding image noise and unmodeled dynamics, variable ξ is very close zero at the end of the ﬁrst step as shown in Fig. 6(a). After the ﬁrst step, the translation discussed in Sect. 3.2 is performed. The robot moves along fc fd and its velocity decreases as the distance si (t) get close to zero (see Fig. 6(b)). The overall performance of the visual servoing strategy was interesting. In fact, the ﬁnal conﬁguration diﬀers from the desired one of only 1.96cm corresponding to an average distance of about 5 pixels between ﬁnal and target feature points. The complete motion of the feature points is shown in Fig. 7. Note that the trajectory discontinuity of the features is due to the change of the control law between the ﬁrst and the second step. Note also that the perfect bi-conic intersection (i.e., equal orientation to the desired position) is lost during the second step (compare Fig. 5(f) with

IBVS for Mobile Robots with Catadioptric Camera

167

desired position

initial position

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 5. Image-based visual servoing: 2-steps algorithm. (a) Initial and desired robot positions; (b) When a rotational disparity occurs then all bi-conic do not intersect; (c) Rotation compensation after the 1st step; (d) After the 1st step all bi-conics intersect at the same two points; (e) Translational motion to the target position; (f) Current feature points (dot) reach the desired ones (cross).

Fig. 5(d)). The main reason of this error is the slippage of the wheels during the pure translation. In order to improve performances and to increase robustness with respect to wheels slippage and other unmodeled dynamics and errors, we implemented an iterative procedure as in [14] to robustly perform the proposed visual servoing procedure. Experiments have been executed also with this iterative implementation and more details are in [11]. Globally, the iterated-2steps algorithm exhibits good performances for the robot motion, and the positioning error between the ﬁnal and the desired camera-robot conﬁgurations decreases from 2cm to 0.5cm (iterative).

168

G.L. Mariottini, D. Prattichizzo, and A. Carbella -6

x 10 4

250

3.5 200

3 s [pixels]

Csi

2.5 2 1.5 1

150

100

50

0.5 0

5

10

Steps

15

20

0

(a)

50

100

150

200 Steps

250

300

350

(b)

Fig. 6. Visual servoing (2-steps). (a) First step: Parameter ξ(t) decreases (rotation compensation); (b) Second step: Distance s(t) between corresponding feature points measured along bi-conics decreases.

Fig. 7. Visual servoing. Motion of the features points in the image plane from the initial (circle) to the desired image (cross).

5 Conclusions and Future Work In this paper we presented an image-based visual servoing for holonomic mobile robot equipped with a catadioptric camera. This vision sensor combines

IBVS for Mobile Robots with Catadioptric Camera

169

lens and mirrors to enlarge the ﬁeld of view. The main feature of our technique consists in exploiting the auto-epipolar property, a special conﬁguration which occurs when the desired and the current views undergo a pure translation. This condition can be detected directly in the image plane, without requiring any estimation and it has been used to compensate the rotation disparity between initial and desired image (1st step). Then a feature-based translational step (2nd step) is needed to drive the robot to the target position. Experimental results on a real robot validate the applicability of our image-based strategy. A current hypothesis of this work is that camera calibration parameters are know. Future work will deal with the extension to uncalibrated catadioptric cameras.

References 1. J.P. Barreto, F. Martin, and R. Horaud. Visual servoing/tracking using central catadioptric images. In 8th International Symposium on Eperimental Robotics, pages 863–869, 2002. 2. F. Conticelli, B. Allotta, and P. K. Khosla. Image-based visual servoing of nonholonomic mobile robots. In 38th IEEE Conference on Decision and Control, 1999. 3. N. Cowan, O. Shakernia, R. Vidal, and S. Sastry. Vision-based follow the leader. In IEEE Intelligient Robots and Systems, pages 1797–1801, 2003. 4. J. Gaspar, N. Winters, and J. Santos-Victor. Vision-based navigation and environmental representations with an omnidirectional camera. In IEEE Transactions on Robotics and Automation, volume 16, pages 890–898, 2000. 5. C. Geyer and K. Daniilidis. A unifying theory for central panoramic systems. In 6th European Conference on Computer Vision, pages 445–462, 2000. 6. R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 7. K. Hashimoto and T. Noritsugu. Visual servoing of nonholonomic cart. In IEEE International Conference on Robotics and Automation, pages 1719–1724, 1997. 8. S. A. Hutchinson, G. D. Hager, and P. I. Corke. A tutorial on visual servo control. IEEE Transaction on Robotics and Automation, 12(5):651–670, 1996. 9. Y. Ma, S. Soatto, J. Koˇseck´ a, and S. Shankar Sastry. An invitation to 3-D Vision, From Images to Geometric Models. Springer Verlag, New York, 2003. 10. E. Malis, F. Chaumette, and S. Boudet. 2-1/2-D visual servoing. IEEE Transactions on Robotics and Automation, 15(2):238–250, 1999. 11. G.L. Mariottini and D. Prattichizzo. Image-based visual servoing for holonomic robots with central catadioptric camera. In Internal Report, http://www.dii.unisi.it/~sirslab/MaPr_IR05.pdf, September 2005. University of Siena. 12. Y. Mezouar, H. Haj Abdelkader, P. Martinet, and F. Chaumette. Central catadioptric visual servoing from 3d straight lines. In 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems, volume 1, pages 343–349, 2004. 13. S.K. Nayar. Catadioptric omnidirectional camera. In International Conference on Computer Vision and Pattern Recognition, pages 482–488, 1997.

170

G.L. Mariottini, D. Prattichizzo, and A. Carbella

14. G. Oriolo and M. Vendittelli. A framework for the stabilization of general nonholonomic systems with an application to the plate-ball mechanism. IEEE Transactions on Robotics, 21(2):162–175, 2005. 15. J. Piazzi, N. Cowan, and D. Prattichizzo. Auto-epipolar visual servoing. In Proc. IEEE Conference on Intelligent Robots and Systems, 2004. 16. R. Pissard-Gibollet and P. Rives. Applying visual servoing techniques to control a mobile hand-eye system. In IEEE Transactions on Robotics and Automation, pages 166–171, 1995. 17. K. Usher, P. Ridley, and P. Corke. Visual servoing for a car-like vehicle - An application of omnidirectional vision. In 2003 IEEE International Conference on Robotics and Automation, pages 4288–4293, 2003.

Multi-knowledge Approach for Mobile Robot Identiﬁcation of a Changing Environment QingXiang Wu12 , T.M. McGinnity1 , Girijesh Prasad1 , David Bell2 and Brendan Glackin1 1 2

School of Computing and Intelligent Systems, University of Ulster at Magee, Londonderry, BT48 7JL, N.Ireland, UK [email protected] Department of Computer Science, Queens University, Belfast, UK [email protected]

Summary. In this paper, a large environment is divided into sub-areas to enable a robot to apply precise localization technology eﬃciently in real time. Sub-area features are represented in a feature information system so that conventional machine learning or data mining approaches can be applied to identify the sub-areas. However, conventional representations with a single body of knowledge encounter many problems when the sub-area features are changed. In order to deal with changing environments, the multi-knowledge approach is applied to the identiﬁcation of environments. Multi-knowledge is extracted from a feature information system by means of multiple reducts (feature sets) so that a robot with multi-knowledge is capable of identifying an environment with some changing features. A case-study demonstrates that a robot with multi-knowledge can cope better with the identiﬁcation of an environment with changing features than conventional single body of knowledge.

1 Introduction Current robot localization techniques include position tracking techniques [1, 2, 3], global localization techniques [4], Markov localization [5, 6, 7], landmark localization [8, 9, 10], and visual-based localization [11, 12, 13, 14]. These techniques aim to estimate the position of a robot within a given environment by means of sensor data and a map of environment. In order to estimate a robot’s position with a high precision, a ﬁne-grained discretization in the state space is usually applied. The spatial resolution is usually between 5 cm and 50 cm and the angular resolution is usually 2 to 5 degrees. For a mediumsized environment of 40 x 40 m2 , an angular grid resolution of 2o , and a cell size of 10 x 10cm2 the state space consists of 28,800,000 states. If a map of an environment is very large, localization techniques encounter a high cost in computation, and are very diﬃcult to use in real time. In this paper, a strategy is proposed to address this problem. A large map may be divided into area

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 171–180, 2006. © Springer-Verlag Berlin Heidelberg 2006

172

Q. Wu et al.

maps, and a robot can apply conventional machine learning approaches to recognize area environment features, for example, in a building, in a room, or in a car park. Therefore, the robot carries out a precise localization in a small area using the speciﬁc information on the area environment such as an electronic map. However, conventional single body of knowledge encounters problems in a changing environment. A novel multi-knowledge approach is proposed to solve this problem. In Section 2, an environment feature decision system is used to divide a large map into sub-maps so that conventional machine learning [15] and data mining methods [16, 17] can be applied to identify the environment. The problems with conventional approaches are analyzed in Section 3. In Section 4 the multi-knowledge representation is deﬁned and it is demonstrated that a robot with multi-knowledge has the ability to identify its environment even if some features of the environment changed. The conclusion is given in Section 5.

2 Representation of Environment Features 2.1 Environment Features In order to illustrate the application of multi-knowledge to robot identiﬁcation of a changing environment, suppose that there is a robot in a building with six rooms shown in Fig. 1. Rooms from No.1 to No.3 are oﬃces. The area marked 4 is the corridor and rooms No.5 and No.6 are laboratories. Features of the environment are selected from those that can be detected by the sensors in the robot. In this case, the features of the environment are Area, GroundColour, WallColour, CeilingHeight, and Illumination. After the robot had gone through all the rooms in the building, the environment feature information system was obtained. The colours are obtained from a colour camera. Illumination is detected by photosensitive sensors. Distances are detected by distance detectors.

Fig. 1. A robot in a building with six rooms

Mobile Robot Identiﬁcation of a Changing Environment

173

Fig. 2. A robot detects a room area

Area is calculated from the distance data. As shown in Fig. 2, let r(θ) represent distance from the robot to a wall along the direction θ. The distance r(θ) can be detected by distance detectors in the robot. The area can be calculated as follows. s=

1 2

2π 0

r2 (θ)dθ

(1)

For simplicity, r(θ) is represented by a set of distances r0 , r1 , r2 , ..., rN with angular interval Δθ = 2π/N . The area between rn and rn+1 can be calculated by using the small triangle if Δθ is very small. 1 rn rn+1 Δθ 2 The room area can be calculated by sn =

N −1

N −1

sn =

s= n=0

n=0

1 rn rn+1 Δθ 2

(2)

(3)

where r 0 = r N = r(θ = 0). Note that the robot can detect the area with this formula at any point in the room. This is an approximate formula. The error depends on the angular interval Δθ is. The other features can be detected by the corresponding sensors in the robot. An example environment information system is shown in Table 1. 2.2 Environment Feature Instance System Following [17], let I =< U, A ∪ D > represent an instance system, where U = u1 , u2 , ... ,ui , ..., u|U | is a ﬁnite non-empty set, called an instance space or universe, and where ui is called an instance in U . A = a1 , a2 , a3 , ..., ai , ..., a|A| , also a ﬁnite non-empty set, is a set of attributes of the instances, where ai

174

Q. Wu et al.

Table 1. Room Feature Decision System S:Area,GC:GroundColor, WC:WallColor, CH:CeilingHeight, B:Illumination, R:Room U S GC

WC

CH B R

1 2 3 4 5 6

yellow white white white yellow white

2.6 2.6 2.6 2.8 3.0 3.0

9 12 9 13 15 15

yellow yellow blue gray gray yellow

1 2 1 2 3 3

1 2 3 4 5 6

is an attribute of a given instance. D is a non-empty set of decision attributes, and A ∩ D = 0. For every a ∈ A there is a domain, represented by Va , and there is a mapping a(u) : U → Va from U into the domain Va , where a(u) represents the value of attribute a of instance u and is a value in the set Va . Va = a(U ) = a(u) : u ∈ U f or a ∈ A

(4)

For a decision system, the domain of decision attribute is represented by Vd = d(U ) = d(u) : u ∈ U f or d ∈ D

(5)

For example, Table 1 is regarded as an environment feature decision system in which the information is detected by a robot that has gone through all the rooms in the building shown in Fig 1. The attribute set A =Area, GroundColor, WallColor, CeilingHeight, Illumination. The decision attribute is D=Room. According to Equation ( 4), VArea = 9, 12, 13, 15. So |VArea | = 4. By analogy, we have |VGround | = 3, |VW all | = 2.|VCeiling | = 2, |VLight | = 3, and the size of decision attribute domain |VRoom | = 6. 2.3 Condition Vector Space The condition vector space, which is generated from attribute domain Va , is denoted by V×A = × Va = Va1 × Va2 × ... × Va|A| a∈A

(6)

and the size of the space is |A|

|Vai |

|V×A | = i=1

(7)

Mobile Robot Identiﬁcation of a Changing Environment

175

For Table 1, |A|

|V×A | =

|Vai | = |VArea | × |VGround | × |VW all | × |VCeiling | × |VLight | (8) i=1

|V×A | = 4 × 3 × 2 × 2 = 48

(9)

This means that there are 48 condition vectors in the condition vector space |V×A |. Every condition vector corresponds to a combination of attribute values. Every instance corresponds to a vector in the vector space by its attribute values. A(u) = (a1 (u), a2 (u), ...., a|A| (u))

(10)

For example, A(Room1) = (9, yellow, yellow, 2.6, 1). If A(u) = A(v) for u ∈ U and v ∈ U , instance u and instance v have the same condition vector. Instance u and instance v are indiscernible. A(U ) represents a set of vectors which exist in the decision system. A(U ) = {A(u) : u ∈ U }

(11)

If |A(U )| = |V×A |, the system is called a complete instance system or complete system. In the real world, training sets for decision-making or feature classiﬁcation are rarely completed systems. Table 1 has only 6 existing condition vectors, while the size of condition space is |V×A | = 48. Thus Table 1 is not a complete instance system. The decision vector space is represented by V×D = × Vd = Vd1 × Vd2 × ... × Vd|D| d∈D

(12)

Vd is equivalent to V×D in cases in which there is only one decision attribute. 2.4 Conventional Knowledge Representations Knowledge can be represented in many forms such as rules, decision trees, neural networks, Bayesian belief networks and other methods [15]. In general, knowledge can be deﬁned as a mapping from condition vector space to decision space. ϕA : V×A → V×D

(13)

V×B = × Va = Va1 × Va2 × ... × Va|B|

(14)

Let B ⊂ A. We have a∈B

and

176

Q. Wu et al. |B|

|Vbi |

|V×B | =

(15)

i=1

|V×B | is called a subspace of |V×A |. The conventional approach in machine learning or data mining is to select a “best” subset B to get a mapping ϕB based on the subset B. ϕB : V×B → V×D

(16)

ϕB is then applied instead of ϕA to make decisions or in feature classiﬁcation. Diﬀerent mappings can be obtained because the training set is an incomplete system. For example, the decision tree, which is shown in Figure 3, is obtained from Table 1 by means of information entropy.

Fig. 3. Knowledge ϕB represented by a decision tree

The advantage of this approach is that it enables a robot to identify the rooms in the building with only two attributes – Area and Ground.

3 Problems with Single Knowledge A single body of knowledge encounters problems in cases of dealing with changing attributes, incomplete attributes, and unseen instances. Issues arise when using a single body of knowledge in a changing environment. For example suppose that the features of the rooms in Table 1 have been changed. The new features are shown in Table 2. Using the decision tree in Fig. 3 consider a robot that goes through the rooms with the changed features as shown in Table 2. When the robot enters room 1, it obtains the condition vector (9, yellow, yellow, 2.6, 2). As the decision tree only tests Area and GroundColor, it does not care that Illumination has changed.

Mobile Robot Identiﬁcation of a Changing Environment

177

Table 2. Changed Room Feature Values S:Area,GC:GroundColor, WC:WallColor, CH:CeilingHeight, B:Illumination, R:Room US

GC

WC

CH

B

R

1 2 3 4 5 6

yellow yellow brown(blue) blue(gray) gray yellow

yellow white white yellow(white) yellow white

2.6 2.8(2.6) 2.9(2.6) 2.8 3.0 3.0

2(1) 2 3(1) 2 3 3

1 2 3 4 5 6

9 12 9 13 14(15) 13(15)

By using the decision tree, the robot can determine that it is in room 1. For the other instances from U2 to U6, the results are listed as follows. U2: (12, yellow, white, 2.8, 2) → decision tree → Room2. U3: (9, black, white, 2.9, 3) → decision tree → Unknown. U4: (13, blue, yellow, 2.8, 2) → decision tree → Room4. U5: (14, gray, yellow, 3.0, 3) → decision tree → Unknown. U6: (13, yellow, white, 3.0, 3) → decision tree → Room4. In many cases, the decision tree will give a wrong answer or cannot give any answer. This problem is not only encountered by decision trees, but other methods as well when using single body representations.

4 Environment Identiﬁcation Based on Multi-knowledge A multi-knowledge approach can be applied in the machine-learning domain to solve diﬀerent problems. For example a combination of multi-knowledge and the Bayes Classiﬁer have been used successfully to improve the accuracy of a classiﬁcation system [18]. Here we apply the concept of multi-knowledge to a robot to identify the changed environment. A single body of knowledge representation usually does not need all attributes in an instance system. For example, the decision tree in Figure 3 contains only two attributes-Area and GroundColor. This attribute set is called a reduct. In general, there are many reducts in a decision system. In a conventional data mining method, ﬁrstly, a good reduct(or a set of good features) is selected and then extracted to form a single body of knowledge based on this reduct. The multi-knowledge approach however encourages ﬁnding as many reducts as possible. Every reduct can contribute to a single body of knowledge. A set of these single bodies of knowledge is called the multi-knowledge. Deﬁnition: Given a decision system I = < U, A ∪ D >. Multi-knowledge is deﬁned as Φ = {ϕB |B ∈ RED}

(17)

178

Q. Wu et al.

where ϕB is a mapping from the condition vector space V×B to the decision space V×D . RED is a set of reducts from the decision system. Reducts RED can be found by the algorithm in [18]. For Table 1, there are 5 reducts. RED={{Area, Wall}, {Area, Ground}, {Ground, Illumination}, {Ground, Ceiling, Wall}, {Wall, Illumination, Ceiling}} Applying these 5 reducts, 5 single bodies of knowledge can be obtained. For example, let reduct B={Area, Wall}. The existing condition vector space is A{Area,W all} (U ) = {(9, yellow), (12, white), (9, white), (13, yellow), (15, yellow), (15, white)}

(18)

Extracting rules in this condition vector space from Table 1 and generalizing the rules, we have 

ϕ(Area,W all)

Room1 Room2  Room3 = Room4 Room5 Room6  U nknown f or other cases

if (Area, W all) = (9, yellow) if (Area) = (12) if (Area, W all) = (9, white) if (Area) = (13) if (Area, W all) = (15, yellow) if (Area, W all) = (15, white)

(19) where ϕ(Area,W all) is presented by using the existing condition vector space A(Area,W all) instead of V×(Area,W all) because the decision system is not a complete system. ϕ(Area,Ground) , ϕ(Ground,Illumination) , ϕ(Ground,Ceiling,W all) and ϕ(Illumination,Ceiling,W all) can be obtained by analogy. Every instance may get multiple decisions from multi-knowledge. In order to merge these decisions, a decision support degree, which is denoted by Sup(di ), is deﬁned by a probability as follows. Sup(di ) = P (di |di = ϕB )

f or ϕB ∈ Φ

(20)

where di ∈ Vd is a decision in the decision space, ϕB is a single body of knowledge among the multi-knowledge Φ. The ﬁnal decision is made by dF inal = arg max Sup(di )

(21)

di ∈Vd

Now consider a robot has multi-knowledge Φ extracted from Table 1 and it enters the changed environment shown in Table 2. For instance U3 in Table 2, ϕ(Area,W all) ϕ(Area,Ground)

ϕ(Ground,Illumination) ϕ(Ground,Ceiling,W all)

ϕ(Illumination,Ceiling,W all)

= ϕ(9,white) = ϕ(9,brown) = ϕ(brown,3) = ϕ(brown,2.9,white) = ϕ(3,2.9,white)

= = = = =

Room3 U nknown U nknown U nknown U nknown

(22)

Mobile Robot Identiﬁcation of a Changing Environment

179

So Sup(Room1) = 0, Sup(room2) = 0, Sup(Room3) = 1, Sup(Room4) = 0, Sup(Room5) = 0, Sup(Room6) = 0. The ﬁnal decision is as follows. dF inal = arg max Sup(di ) = Room3

(23)

di ∈Vd

It is possible, according to this algorithm, to calculate dF inal for all the instances in Table 2, The results are shown in Table 3. The results show that multi-knowledge can cope with a changing environment much better than any single body of knowledge. Table 3. Results for multi-knowledge to identify the changed environment Φ

U1

U2

U3

U4

U5

U6

ϕ(Area,W all) ϕ(Area,Ground) ϕ(Ground,Illum)

Room1 Room1 Room2 Room1 Unknown Room1

Room2 Room2 Room2 Unknown Unknown Room2

Room3 Unknown Unknown Unknown Unknown Room3

Room4 Room4 Unknown Unknown Unknown Room4

Unknown Unknown Room5 Room5 Room5 Room5

Unknown Unknown Room6 Room6 Room6 Room6

ϕ(Ground,Ceiling,W all) ϕ(Illum,Ceiling,W all) dF inal

5 Conclusion In this paper the problem of a robot identifying its location in a changing environment has been considered. A large environment is divided into small area environments. An instance information system is applied to represent environment features of the sub area environments. Conventional machine leaning approaches are applied to identify sub area environments. The problem for identifying a changing environment was analyzed, where a robot using conventional machine leaning mechanisms ﬁnds it diﬃcult to solve the problem. In contrast a multi-knowledge approach was proposed to solve the problem. A case-study was presented to demonstrate that a robot with multi-knowledge copes with a changing environment much better than a conventional singlebody knowledge representation.

References 1. Leonard JJ, Durrant-Whyte HF (1991) Mobile robot localization by tracking geometric beacons, IEEE Transactions on Robotics and Automation, 7(3): 376−382. 2. Feng L, Borenstein J, Everett HR (1994) ”Where am I?” Sensors and methods for autonomous mobile robot positioning. Technical Report UMMEAM9412, University of Michigan, Ann Arbor, MI.

180

Q. Wu et al.

3. Jensfelt P, Kristensen S (2001) Active global localization for a mobile robot using multiple hypothesis tracking, IEEE Transactions on Robotics and Automation, Volume: 17(5): 748−760. 4. Dissanayake MW, Newman MG, Clark S, Durrant-Whyte, HF, Csorba M (2001) A solution to the simultaneous localization and map building (SLAM) problem. IEEE Transactions on Robotics and Automation, 17(3): 229−241. 5. Fox D, Burgard W, and Thrun S (1999) Markov Localization For Mobile Robots in Dynamic Environments. Journal of Artiﬁcial Intelligence Research,11: 391−427. 6. Fox D, Burgard W, Kruppa H, and Thrun S (2000) A probabilistic approach to collaborative multi-robot localization. Autonomous Robots, Special Issue on Heterogeneous Multi-Robot Systems, 8(3): 325−344. 7. Thrun S (2001) A probabilistic online mapping algorithm for teams of mobile robots. International Journal of Robotics Research, 20(5): 335−363. 8. Thrun S, (1998) Bayesian Landmark Learning for Mobile Robot Localization. Machine Learning, 33(1): 41−76. 9. Shimshoni I (2002) On mobile robot localization from landmark bearings, IEEE Transactions on Robotics and Automation, V18(6):971−976. 10. Briechle K, Hanebeck UD (2004), Localization of a mobile robot using relative bearing measurements; IEEE Transactions on Robotics and Automation, 20(1): 36−44. 11. Se S, Lowe D, Little J (2001) Vision-based mobile robot localization and mapping using scale-invariant features. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pp.2051-2058, Seoul, Korea. 12. Niall W, Jos SV (2002) Information Sampling for vision-based robot navigation, Robotics and Autonomous Systems, 41( 2-3): 145−159. 13. Ayache N (1991) Artiﬁcial Vision for Mobile robots - Stereo-vision and Multisensor Perception, MIT-Press, page 342. 14. Kiyosumi K, Jun M, Yoshiaki S (2002) Autonomous visual navigation of a mobile robot using a human-guided experience, Robotics and Autonomous Systems, 40(2-3): 121−130. 15. Mitchell MT (1997) Machine Learning, McGraw Hill, Co-published by the MIT Press Companies, Inc. 16. Lin TY, Cercone N(eds) (1997) Rough Sets and Data Mining: Analysis for Imprecise Data, Boston, Mass; London : Kluwer Academic. 17. Polkowski L, Tsumoto S, Lin TY (2000) Rough Set Methods and Applications, New Developments in Knowledge Discovery in Information Systems, PhysicaVerlag, A Springer-Verlag Company. 18. Wu QX, Bell DA, McGinnity TM (2005) Multi-knowledge for Decision Making, International Journal of Knowledge and Information Systems, Springer-Verlag London Ltd. 7(2): 246−266.

Robust Monte-Carlo Localization Using Adaptive Likelihood Models Patrick Pfaﬀ1 , Wolfram Burgard1 , and Dieter Fox2 1 2

Department of Computer Science, University of Freiburg, Germany, {pfaff,burgard}@informatik.uni-freiburg.de Department of Computer Science & Engineering, University of Washington, USA, [email protected]

Summary. In probabilistic mobile robot localization, the development of the sensor model plays a crucial role as it directly inﬂuences the eﬃciency and the robustness of the localization process. Sensor models developed for particle ﬁlters compute the likelihood of a sensor measurement by assuming that one of the particles accurately represents the true location of the robot. In practice, however, this assumption is often strongly violated, especially when using small sample sets or during global localization. In this paper we introduce a novel, adaptive sensor model that explicitly takes the limited representational power of particle ﬁlters into account. As a result, our approach uses smooth likelihood functions during global localization and more peaked functions during position tracking. Experiments show that our technique signiﬁcantly outperforms existing, static sensor models.

1 Introduction Probabilistic mobile robot localization estimates a robot’s pose using a map of the environment, information about the robot’s motion, and sensor measurements that relate the robot’s pose to objects in the map. The sensor model plays a crucial role as it directly inﬂuences the eﬃciency and the robustness of the localization process. It speciﬁes how to compute the likelihood p(z | x, m) or short p(z | x) of a measurement z given the vehicle is at position x in the environment m. A proper likelihood function has to take various sources of noise into account, including sensor uncertainty, inaccuracy of the map, and uncertainty in the robot’s location. An improperly designed likelihood function can make the vehicle overly conﬁdent in its position and in this way might lead to a divergence of the ﬁlter. On the other hand, the model might be deﬁned in a too conservative fashion and this way prevent the robot from localizing as the sensor information cannot compensate the uncertainty introduced by the motion of the vehicle. Monte Carlo localization (MCL) is a particle-ﬁlter based implementation of recursive Bayesian ﬁltering for robot localization [2, 6]. In each iteration of

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 181–194, 2006. © Springer-Verlag Berlin Heidelberg 2006

182

P. Pfaﬀ, W. Burgard, and D. Fox

MCL, the likelihood function p(z | x) is evaluated at sample points that are randomly distributed according to the posterior estimate of the robot location. Sensor models developed for MCL typically assume that the vehicle location x is known exactly; that is, they assume that one of the sampling points corresponds to the true location of the robot. However, this assumption is only valid in the limit of inﬁnitely many particles. Otherwise, the probability that the pose of a particle exactly corresponds to the true location of the robot is virtually zero. As a consequence, these likelihood functions do not adequately model the uncertainty due to the ﬁnite, sample-based representation of the posterior. In practice, a typical solution to this problem is to artiﬁcially inﬂate the noise of the sensor model. Such a solution is not satisfying, however, since the additional uncertainty due to the sample-based representation is not known in advance; it strongly varies with the number of samples and the uncertainty in the location estimate. In this paper we introduce a novel, adaptive sensor model that explicitly takes location uncertainty due to the sample-based representation into account. In order to compute an estimate of this uncertainty, our approach borrows ideas from techniques developed for density estimation. The goal of density estimation is to extract an estimate of the probability density underlying a set of samples. Most approaches to density estimation estimate the density at a point x by considering a region around x, where the size of the region typically depends on the number and local density of the samples (for instance, see Parzen window or k-nearest neighbor approaches [4]). The density estimation view suggests a way for computing the additional uncertainty that is due to the sample-based representation: When evaluating the likelihood function at a sample location, we consider the region a density estimation technique would take into account when estimating the density at that location. The likelihood function applied to the sample is then computed by integrating the point-wise likelihood over the corresponding region. As a result, the likelihood function automatically adapts to the local density of samples, being smooth during global localization and highly “peaked” during position tracking. Our approach raises two questions. 1. How large should the region considered for a sample be? 2. How can we eﬃciently determine this region and integrate the observation likelihood over it? While our idea can be applied to arbitrary particle ﬁlter approaches, this paper focuses on how to address these questions in the context of mobile robot localization. In particular, we estimate the region associated to a particle using a measure applied in k-nearest neighbor density estimation, in which the region of a point grows until a suﬃcient number of particles lies within it. We show that by considering the simple case of k = 1 and learning the appropriate smoothness of the likelihood function, we can eﬀectively improve the speed required for global localization and at the same time achieve high accuracy during position tracking.

Robust Monte-Carlo Localization Using Adaptive Likelihood Models

183

This paper is organized as follows. After discussing related work in the next section, we brieﬂy describe Monte Carlo localization in Section 3. Our approach to dynamically adapt the likelihood model is presented in Section 4. Finally, in Section 5, we present experimental results illustrating the advantages of our model.

2 Related Work In the literature, various techniques for computing the likelihood function for Monte Carlo localization can be found. They depend mainly on the sensors used for localization and the underlying representation of the map m. For example, several authors used features extracted from camera images to calculate the likelihood of observations. Typical features are average grey values [2], color values [8], color transitions [10], feature histograms [17], or speciﬁc objects in the environment of the robot [9, 12, 13]. Additionally, several likelihood models for Monte-Carlo localization with proximity sensors have been introduced [1, 15]. These approaches either approximate the physical behavior of the sensor [7] or try to provide smooth likelihood models to increase the robustness of the localization process [14]. Whereas these likelihood functions allow to reliably localize a mobile robot in its environment, they have the major disadvantage that they are static and basically independent of the state the localization process has. In the past, is has been observed that the likelihood function can have a major inﬂuence on the performance of Monte-Carlo Localization. For example, Pitt and Shepard [11] as well as Thrun et al. [16] observed that more particles are required if the likelihood function is peaked. In the limit, i.e., for a perfect sensor, the number of required particles becomes inﬁnite. To deal with this problem, Lenser and Veloso [9] and Thrun et al. [16] introduced techniques to directly sample from the observation model and in this way ensure that there is a critical mass of samples at the important parts of the state space. Unfortunately, this approach depends on the ability to sample from observations, which can often only be done in an approximate, inaccurate way. Alternatively, Pitt and Shepard [11] apply a Kalman ﬁlter lookahead step for each particle in order to generate a better proposal distribution. While this technique yields superior results for highly accurate sensors, it is still limited in that the particles are updated independently of each other. Hence, the likelihood function does not consider the number and density of particles. Another way of dealing with the limitations of the sample-based representation is to dynamically adapt the number of particles, as done in KLD sampling [5]. However, for highly accurate sensors, even such an adaptive technique might require a huge number of samples in order to achieve a suﬃciently high particle density during global localization. The reader may notice that Kalman ﬁlter based techniques do consider the location uncertainty when integrating a sensor measurement. This is done

184

P. Pfaﬀ, W. Burgard, and D. Fox

by mapping the state uncertainty into observation space via a linear approximation. However, Kalman ﬁlters have limitations in highly non-linear and multi-modal systems, and the focus of this paper is to add such a capability to particle ﬁlters. More speciﬁcally, in this paper we propose an approach that dynamically adapts the likelihood function during localization. We dynamically calculate for each particle the variance of the Gaussian governing the likelihood function depending on the area covered by that particle.

3 Monte Carlo Localization To estimate the pose x of the robot in its environment, we consider probabilistic localization, which follows the recursive Bayesian ﬁltering scheme. The key idea of this approach is to maintain a probability density p(x) of the robot’s own location, which is updated according to the following rule: p(xt | z1:t , u0:t−1 ) = α · p(zt | xt ) ·

p(xt | ut−1 , xt−1 ) · p(xt−1 ) dxt−1 (1)

Here α is a normalization constant ensuring that p(xt | z1:t , u0:t−1 ) sums up to one over all xt . The term p(xt | ut−1 , xt−1 ) describes the probability that the robot is at position xt given it executed the movement ut−1 at position xt−1 . Furthermore, the quantity p(zt | xt ) denotes the probability of making observation zt given the robot’s current location is xt . The appropriate computation of this quantity is the subject of this paper. Throughout this paper we use a sample-based implementation of this ﬁltering scheme also denoted as Monte Carlo localization [2, 6]. In Monte-Carlo localization, which is a variant of particle ﬁltering [3], the update of the belief generally is realized by the following two alternating steps: 1. In the prediction step, we draw for each sample a new sample according to the weight of the sample and according to the model p(xt | ut−1 , xt−1 ) of the robot’s dynamics given the action ut−1 executed since the previous update. 2. In the correction step, the new observation zt is integrated into the sample set. This is done by bootstrap re-sampling, where each sample is weighted according to the likelihood p(zt | xt ) of sensing zt given by its position xt .

4 Dynamic Sensor Models The likelihood model p(z | x) plays a crucial role in the correction step of the particle ﬁlter. Typically, very peaked models require a huge number of particles. This is because even when the particles populate the state space very densely, the likelihoods of an observation might diﬀer by several orders of

Robust Monte-Carlo Localization Using Adaptive Likelihood Models

185

magnitude. As the particles are drawn proportional to the importance weights, which themselves are calculated as the likelihood of zt given the pose xt of the corresponding particle, a minor diﬀerence in xt will result in a diﬀerence of one order of magnitude in the likelihood and thus will result in a depletion of such a particle in the re-sampling step. Accordingly, an extremely high density of particles is needed when the sensor is highly accurate. On the other hand, the sheer size of the state space prevents us from using a suﬃciently large number particles during global localization in the case that the sensor is highly accurate. Accordingly the sensor model needs to be less peaked in the case of global localization, when the particles are sparsely distributed over the state space. This is essentially the idea of the approach proposed in this paper. 4.1 Likelihood Model for Range Sensors The likelihood model used throughout this paper calculates p(z | x) based on the distance d to the closest obstacle in the direction of the measurement given x (and the map m). Accordingly, p(z | x) is calculated as p(z | x) = p(z | d).

(2)

To determine this quantity, we follow the approach described in Thrun et al. [15] and Choset et al. [1]. The key idea of this sensor model is to use a mixture of four diﬀerent distributions to capture the noise and error characteristics of range sensors. It uses a Gaussian phit (z | d) to model the likelihood in situations in which the beam hits the next object in the direction of the measurement. Additionally, it uses a uniform distribution prand (z | d) to represent random measurements. It furthermore models objects not contained in the map as well as the eﬀects of cross-talk between diﬀerent sensors by an exponential distribution pshort (z | d). Finally, it deals with detection errors, in which the sensor reports a maximum range measurement, using a uniform distribution pmax (z | d). These four diﬀerent distributions are mixed by a weighted average, deﬁned by the parameters αhit , αshort , αmax , and αrand . p(z | d)

(3) T

= (αhit , αshort , αmax , αrand ) · (phit (z | d), pshort (z | d), pmax (z | d), prand (z | d)) . Note that these parameters need to satisfy the constraints that none of them should be smaller than zero and that p(z | d) integrates to 1 over all z for a ﬁxed d. A plot of this sensor model for a given set of parameters is shown in Figure 1. Also note that the exponential distribution is only used to model measurements that are shorter than expected, i.e., for measurements z with z < d. Therefore, there is a discontinuity at z = d (see Thrun et al. [15] for details).

186

P. Pfaﬀ, W. Burgard, and D. Fox sigma_von_r

sigma_2

sigma_1

radius_r_mal_alpha

Fig. 1. Sensor model given by a mixture of diﬀerent distributions (left image) and piecewise linear function used to calculate the standard deviation based on the distance d = 2r to the closest particle.

4.2 Adapting the Variance As already mentioned above, the particle set should approximate the true posterior as closely as possible. Since we only have a limited number of particles, which in practice is often chosen as small as possible to minimize the computational demands, we need to take into account potential approximation errors. The key idea of our approach is to adjust the variance of the Gaussian governing phit (z | d), which models the likelihood of measuring z given that the sensor detects the closest obstacle in the map, such that the particle set yields an accurate approximation of the true posterior. To achieve this, we approximatively calculate for each particle i the space it covers by adopting the measure used in k-nearest neighbor density estimation [4]. For eﬃciency reasons we rely on the case of k = 1, in which the spatial region covered by a particle is given by the minimum circle that is centered at the particle and contains at least one neighboring particle in the current set. To calculate the radius ri of this circle, we have to take both, the Euclidean distance of the positions and the angular diﬀerence in orientation, into account. In our current implementation we calculate ri based on a weighted mixture of the Euclidean distance and the angular diﬀerence d(x, x ) =

(1 − ξ)((x1 − x1 )2 + (x2 − x2 )2 ) + ξ δ(x3 − x3 )2 ,

(4)

where x1 and x1 are the x-coordinates, x2 and x2 are the y-coordinates, and δ(x3 − x3 ) is the diﬀerences in the angle of x and x . Additionally, ξ is a weighing factor that was set to 0.8 in all our experiments. Figure 2 shows the fraction of a map and a particle distribution. The circle around each particle represents the radius r = 12 d(x, x ) to the closest particle x . The next step is to adjust for each particle the standard deviation σ of the Gaussian in phit (z | d) based on the distance r = 12 d(x, x∗ ), where x∗ is the particle closest to x with respect to Equation 4. In our current implementation we use a piecewise linear function σ(r) to compute the standard deviation of phit (z | d):

Robust Monte-Carlo Localization Using Adaptive Likelihood Models

187

Fig. 2. Fraction of a particle set consisting of 10,000 particles during a global localization run. The circles around the individual particles represent the radius r = 21 d(x, x ) calculated from the distance to the particle x closest to x.

Fig. 3. Distribution of 1,000,000 particles after 0, 1 and 2 integrations of measurements with the sensor model according to the speciﬁcations of the SICK LMS sensor.

  σ1 σ(r) = σ2  αr

if αr < σ1 if αr > σ2 otherwise.

(5)

To learn the values of the parameters σ1 , σ2 , and α of this function, we performed a series of localization experiments on recorded data, in which we determined the optimal values by minimizing the average distance of the obtained distributions from the true posterior. Since the true posterior is unknown, we determined a close approximation of it by increasing the number of particles until the entropy of a three-dimensional histogram calculated from the particles did not change anymore. In our experiment this turned out to be the case for 1,000,000 particles. Throughout this experiment, the sensor model corresponded to the error values provided in the speciﬁcations of the laser range ﬁnder used for localization. In the remainder of this section, we will denote the particle set representing the true posterior by S ∗ . Figure 3 shows the set S ∗ at diﬀerent points in time for the data set considered in this paper.

188

P. Pfaﬀ, W. Burgard, and D. Fox

To calculate the deviation of the current particle set S from the true posterior represented by S ∗ , we evaluate the KL-distance between the distributions obtained by computing a histogram from S and S ∗ . Whereas the spatial resolution of this histogram is 40cm, the angular resolution is 5 degrees. For discrete probability distributions, p = p1 , . . . , pn and q = q1 , . . . , qn , the KLdistance is deﬁned as KL(p, q) =

pi log2 i

pi . qi

(6)

Whenever we choose a new standard deviation for phit (z | d), we adapt the weights αhit , αshort , αmax , and αrand to ensure that the integral of the resulting density is 1. Note that in principle phit (z | d) should encode also several other aspects to allow for a robust localization. One such aspect, for example, is the dependency between the individual measurements. For example, a SICK LMS laser range scanner typically provides 361 individual distance measurements. Since a particle never corresponds to the true location of the vehicle, the error in the pose of the particle renders the individual measurements as dependent. This dependency, for example, should also be encoded in a sensor model to avoid the ﬁlter becoming overly conﬁdent about the pose of the robot. To reduce potential eﬀects of the dependency between the individual beams of a scan, we only used 10 beams at angular displacements of 18 degrees from each scan.

5 Experimental Results The sensor model described above has been implemented and evaluated using real data acquired with a Pioneer PII DX8+ robot equipped with a laser range scanner in a typical oﬃce environment. The experiments described in this section are designed to investigate if our dynamic sensor model outperforms static models. Throughout the experiments we compared our dynamic sensor model with diﬀerent types of alternative sensor models: 1. Best static sensor model. This model has been obtained by evaluating a series of global localization experiments, in which we determined the optimal variance of the Gaussian by maximizing the utility function I

U (I, N ) =

(I − i + 1) i=1

Pi , N

(7)

where I is the number of integrations of measurements into the belief during the individual experiment, N is the number of particles, and Pi is the number of particles lying in a 1m range around the true position of the robot.

Robust Monte-Carlo Localization Using Adaptive Likelihood Models

189

2. Best tracking model. We determined this model in the same way as the best static sensor model. The only diﬀerence is that we have learned it from situations in which the ﬁlter was tracking the pose of the vehicle. 3. SICK LMS model. This model has been obtained from the hardware speciﬁcations of the laser range scanner. 4. Uniform dynamic model. In our dynamic sensor model the standard deviation of the likelihood function is computed on a per-particle basis. We also analyzed the performance of a model in which a uniform standard deviation is used for all particles. The corresponding value is computed by taking the average of the individual standard deviations.

Fig. 4. Distribution of 10,000 particles after 1, 2, 3, 5, and 11 integrations of measurements with our dynamic sensor model (left column) and with the best static sensor model (right column).

5.1 Global Localization Experiments The ﬁrst set of experiments is designed to evaluate the performance of our dynamic sensor model in the context of a global localization task. In the particular data set used to carry out the experiment, the robot started in the kitchen of our laboratory environment (lower left room in the maps depicted in Figure 3). The evolution of a set of 10,000 particles during a typical global localization run with our dynamically adapted likelihood model and with the

190

P. Pfaﬀ, W. Burgard, and D. Fox

dynamic sensor model uniform dynamic sensor model best static sensor model

100

particles closer than 1m to ground thruth [%]

particles closer than 1m to ground thruth [%]

Fig. 5. Evolution of the distance d(x, x ) introduced in Equation (4). Distribution of 10,000 particles after 1, 3, 4 and 5 integrations of measurements with our dynamic sensor model.

80 60 40 20 0

0

5

10

15 iteration

20

25

dynamic sensor model uniform dynamic sensor model best static sensor model

100 80 60 40 20 0

0

5

10

15

20

25

iteration

Fig. 6. Percentage of particles within a 1m circle around the true position of the robot with our dynamic sensor model, the uniform dynamic model, and the best static model. The left image shows the evolution depending on the number of iterations for 10,000 particles. The right plot was obtained with 2,500 particles.

best static sensor model is depicted in Figure 4. As can be seen, our dynamic model improves the convergence rate as the particles quickly focus on the true pose of the robot. Due to the dynamic adaptation of the variance, the likelihood function becomes more peaked so that unlikely particles are depleted earlier. Figure 5 shows the evolution of the distance d(x, x ) introduced in Equation (4). The individual images illustrate the distribution of 10,000 particles after 1, 2, 3, and 5 integrations of measurements with our dynamic sensor model. The circle around each particle represents the distance r = 21 d(x, x ) to the next particle x . Figure 6 shows the convergence of the particles to the true position of the robot. Whereas the x-axis corresponds to the time step, the y-axis shows the number of particles in percent that are closer than 1m to the true position. Shown are the evolutions of these numbers for our dynamic sensor model,

Robust Monte-Carlo Localization Using Adaptive Likelihood Models dynamic sensor model best static sensor model best tracking model SICK LMS model dynamic sensor model with one beam SICK LMS model with one beam

125

100 success rate [%]

191

75

50

25

0

2500

5000

7500

10000

number of particles

Fig. 7. Success rates for diﬀerent types of sensor models and sizes of particle sets. The shown results are computed in extensive global localization experiments.

a uniform dynamic model, and the best ﬁxed model for 10,000 and 2,500 particles. Note that the best static model does not reach 100%. This is due to the fact that the static sensor model relies on a highly smoothed likelihood function, which is good for global localization but does not achieve maximal accuracy during tracking. In the case of 10,000 particles, the variances in the distance between the individual particles are typically so small, that the uniform model achieves a similar performance. Still, a t-test showed that both models are signiﬁcantly better than the best ﬁxed model. In the case of 2,500 particles, however, the model that adjusts the variance on a per-particle basis performs better than the uniform model. Here, the diﬀerences are signiﬁcant whenever they exceed 20. Figure 7 shows the number of successful localizations after 35 integrations of measurements for a variety of sensor models and for diﬀerent numbers of particles. Here, we assumed that the localization was achieved when the mean of the particles diﬀered by at most 30cm from the true location of the robot. First it shows that our dynamic model achieves the same performance as the best static model for global localization. Additionally, the ﬁgure shows that the static model that yields the best tracking performance has a substantially smaller success rate. Additionally, we evaluated a model, denoted as the SICK LMS model, in which the standard deviation was chosen according to the speciﬁcations of the SICK LMS sensor, i.e., under the assumption that the particles in fact represent the true position of the vehicle. As can be seen, this model yields the worst performance. Furthermore, we evaluated, how the models perform when only one beam is used per range scan. With this experiment we wanted to analyze whether the dynamic model also yields better results in situations in which there is no dependency between the individual beams of a scan. Again, the plots show that our sensor model outperforms

192

P. Pfaﬀ, W. Burgard, and D. Fox

average standard deviation

0.45

10000 particles 7500 particles 5000 particles 2500 particles

0.4

0.35

0.3

0.25

5

10

15

iteration

Fig. 8. Evolution of the average standard deviation during global localization with diﬀerent numbers of particles and our dynamic sensor model. 0.2

0.16

0.16

0.14

0.14

0.12 0.1 0.08

0.12 0.1 0.08

0.06

0.06

0.04

0.04

0.02

5

10

15 iterations

20

dynamic sensor model static sensor model

0.18

localization error

localization error

0.2

dynamic sensor model static sensor model

0.18

25

0.02

5

10

15 iterations

20

25

Fig. 9. Average localization error of 10,000 (left image) and 2,500 (right image) particles during a position tracking task. Our dynamic sensor model shows a similar performance as the best tracking model.

the model, for which the standard deviation corresponds to the measuring accuracy of the SICK LMS scanner. Finally, Figure 8 plots the evolution of the average standard deviations of several global localization experiments with diﬀerent numbers of particles. As can be seen from the ﬁgure, our approach uses more smoothed likelihood functions when operating with few particles (2,500). The more particles are used in the ﬁlter, the faster the standard deviation converges to the minimum value. 5.2 Tracking Performance We also carried out experiments, in which we analyzed the accuracy of our model when the system is tracking the pose of the vehicle. We compared our

Robust Monte-Carlo Localization Using Adaptive Likelihood Models

193

dynamic sensor model to the best tracking model and evaluated the average localization error of the individual particles. Figure 9 depicts the average localization error for two tracking experiments with 10,000 and 2,500 particles. As can be seen from the ﬁgure, our dynamic model shows the same performance as the tracking model whose parameters have been optimized for minimum localization error. This shows, that our dynamic sensor model yields faster convergence rates in the context of global localization and at the same time achieves the best possible tracking performance.

6 Conclusions In this paper we presented a new approach for dynamically adapting the sensor model for particle ﬁlter based implementations of probabilistic localization. Our approach learns a function that outputs the appropriate variance for each particle based on the estimated area in the state space represented by this particle. To estimate the size of this area, we adopt a measure developed for density estimation. The approach has been implemented and evaluated in extensive experiments using laser range data acquired with a real robot in a typical oﬃce environment. The results demonstrate that our sensor model signiﬁcantly outperforms static and optimized sensor models with respect to robustness and eﬃciency of the global localization process.

Acknowledgment The authors would like to thank Dirk Zitterell for fruitful discussions on early versions of this work. This work has partly been supported by the German Research Foundation (DFG) within the Research Training Group 1103 and under research grant SFB/TR-8, and by the National Science Foundation under grant number IIS-0093406.

References 1. H. Choset, K.M. Lynch, S. Hutchinson, G. Kantor, W. Burgard, Kavraki L.E., and S. Thrun. Principles of Robot Motion Planing. MIT-Press, 2005. 2. F. Dellaert, D. Fox, W. Burgard, and S. Thrun. Monte Carlo localization for mobile robots. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 1999. 3. A. Doucet, N. de Freitas, and N. Gordon, editors. Sequential Monte Carlo in Practice. Springer-Verlag, New York, 2001. 4. R. Duda, P. Hart, and D. Stork. Pattern Classiﬁcation. Wiley-Interscience, 2001.

194

P. Pfaﬀ, W. Burgard, and D. Fox

5. D. Fox. Adapting the sample size in particle ﬁlters through KLD-sampling. International Journal of Robotics Research, 22(12):985 – 1003, 2003. 6. D. Fox, W. Burgard, F. Dellaert, and S. Thrun. Monte Carlo localization: Eﬃcient position estimation for mobile robots. In Proc. of the National Conference on Artiﬁcial Intelligence (AAAI), 1999. 7. D. Fox, W. Burgard, and S. Thrun. Markov localization for mobile robots in dynamic environments. Journal of Artiﬁcial Intelligence Research, 11, 1999. 8. H.-M. Gross, A. K¨ onig, Ch. Schr¨ oter, and H.-J. B¨ ohme. Omnivision-based probalistic self-localization for a mobile shopping assistant continued. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), pages 1505–1511, 2003. 9. S. Lenser and M. Veloso. Sensor resetting localization for poorly modelled mobile robots. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2000. 10. E. Menegatti, A. Pretto, and E. Pagello. Testing omnidirectional visionbased Monte-Carlo localization under occlusion. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), pages 2487–2494, 2004. 11. M. K. Pitt and N. Shephard. Filtering via simulation: auxiliary particle ﬁlters. Journal of the American Statistical Association, 94(446), 1999. 12. T. R¨ ofer and M. J¨ ungel. Vision-based fast and reactive Monte-Carlo localization. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), pages 856– 861, 2003. 13. M. Sridharan, G. Kuhlmann, and P. Stone. Practical vision-based Monte Carlo localization on a legged robot. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2005. 14. S. Thrun. A probabilistic online mapping algorithm for teams of mobile robots. International Journal of Robotics Research, 20(5):335–363, 2001. 15. S. Thrun, W. Burgard, and D. Fox. Probabilistic Robotics. MIT-Press, 2005. 16. S. Thrun, D. Fox, W. Burgard, and F. Dellaert. Robust Monte Carlo localization for mobile robots. Artiﬁcial Intelligence, 128(1-2), 2001. 17. J. Wolf, W. Burgard, and H. Burkhardt. Robust vision-based localization by combining an image retrieval system with Monte Carlo localization. IEEE Transactions on Robotics, 21(2):208–216, 2005.

Metric Localization with Scale-Invariant Visual Features Using a Single Perspective Camera Maren Bennewitz, Cyrill Stachniss, Wolfram Burgard, and Sven Behnke University of Freiburg, Computer Science Institute, D-79110 Freiburg, Germany

Abstract. The Scale Invariant Feature Transform (SIFT) has become a popular feature extractor for vision-based applications. It has been successfully applied to metric localization and mapping using stereo vision and omnivision. In this paper, we present an approach to Monte-Carlo localization using SIFT features for mobile robots equipped with a single perspective camera. First, we acquire a 2D grid map of the environment that contains the visual features. To come up with a compact environmental model, we appropriately down-sample the number of features in the ﬁnal map. During localization, we cluster close-by particles and estimate for each cluster the set of potentially visible features in the map using ray-casting. These relevant map features are then compared to the features extracted from the current image. The observation model used to evaluate the individual particles considers the diﬀerence between the measured and the expected angle of similar features. In real-world experiments, we demonstrate that our technique is able to accurately track the position of a mobile robot. Moreover, we present experiments illustrating that a robot equipped with a diﬀerent type of camera can use the same map of SIFT features for localization.

1 Introduction Self-localization is one of the fundamental problems in mobile robotics. The topic was studied intensively in the past. Many approaches exist that use distance information provided by a proximity sensor for localizing a robot in the environment. However, for some types of robots, proximity sensors are not the appropriate choice because they do not agree with their design principle. Humanoid robots, for example, which are constructed to resemble a human, are typically equipped with vision sensors and lack proximity sensors like laser scanners. Therefore, it is natural to equip these robots with the ability of vision-based localization.

H.I. Christensen (Ed.): European Robotics Symposium 2006, STAR 22, pp. 195–209, 2006. © Springer-Verlag Berlin Heidelberg 2006

196

M. Bennewitz et al.

In this paper, we present an approach to vision-based mobile robot localization that uses a single perspective camera. We apply the well-known Monte-Carlo localization (MCL) technique [5] to estimate the robot’s position. MCL uses a set of random samples, also called particles, to represent the belief of the robot about its pose. To locate features in the camera images, we use the Scale Invariant Feature Transform (SIFT) developed by Lowe [15]. SIFT features are invariant to image translation, scale, and rotation. Additionally, they are partially invariant to illumination changes and aﬃne or 3D projection. These properties make SIFT features particularly suitable for mobile robots since, as the robots move around, they typically observe landmarks from diﬀerent angles and distances, and with a diﬀerent illumination. Whereas existing systems, that perform metric localization and mapping using SIFT features, apply stereo vision in order to compute the 3D position of the features [20, 7, 21, 2], we rely on a single camera only during localization. Since we want to concentrate on the localization aspect, we facilitate the map acquisition process by using a robot equipped with a camera and a proximity sensor. During mapping, we create a 2D grid model of the environment. In each cell of the grid, we store those features that are supposed to be at that 2D grid position. Since the number of observed SIFT features is typically high, we appropriately down-sample the number of features in the ﬁnal map. During MCL, we then rely on a single perspective camera and do not use any proximity information. Our approach estimates for clusters of particles the set of potentially visible features using ray-casting on the 2D grid. We then compare those features to the features extracted from the current image. In the observation model of the particle ﬁlter, we consider the diﬀerence between the measured and the expected angle of similar features. By applying the ray-casting technique, we avoid comparing the features extracted out of the current image to the whole database of features (as the above mentioned approaches do), which can lead to serious errors in the data association. As we demonstrate in practical experiments with a mobile robot in an oﬃce environment, our technique is able to reliably track the position of the robot. We also present experiments illustrating that the same map of SIFT features can be used for self-localization by diﬀerent types of robots equipped with a single camera only and without proximity sensors. This paper is organized as follows. After discussing related work in the following section, we describe the Monte-Carlo localization technique that is applied to estimate the robot’s position. In Section 4, we explain how we acquire 2D grid maps of SIFT features. In Section 5, we present the observation model used for MCL. Finally, in Section 6, we show experimental results illustrating the accuracy of our approach to estimate the robot’s position.

Metric Localization with SIFT Features Using a Single Camera

197

2 Related Work Monte-Carlo methods are widely used for vision-based localization and have been shown to yield quite robust estimates of the robot’s position. Several localization approaches are image-based, which means that they store a set of reference images taken at various locations that are used for localization. Some of the image-based methods rely on an omnidirectional camera in order to localize a mobile robot. The advantages of omnidirectional images are the circular ﬁeld of view and thus, the knowledge about the appearance of the environment in all possible gaze directions. Recent techniques were for example presented by Andreasson et al. [1] who developed a method to match SIFT features extracted from local interest points in panoramic images, by Menegatti et al. [16] who use Fourier coeﬃcients for feature matching in omnidirectional images, and by Gross et al. [9] who compare the panoramic images using color histograms. Wolf et al. [23] apply a combination of MCL and an image retrieval system in order to localize a robot equipped with a perspective camera. The systems presented by Ledwich and Williams [12] and by K˘os´ecka and Li [11] perform Markov localization within a topological map. They use the SIFT feature descriptor to match the current view to the reference images. Whenever using those image-based methods, care has to be taken in deciding at which positions to collect the reference images in order to ensure a complete coverage of the space the robot can travel in. In contrast to this, our approach stores features at the positions where they are located in the environment and not for all possible poses the robot can be in. Additionally, localization techniques have been presented that use a database of observed visual landmarks. SIFT features have become very popular for metric localization as well as for SLAM (simultaneous localization and mapping, [21, 2]). Se et al. [20] were the ﬁrst who performed localization using SIFT features in a restricted area. They did not apply a technique to track the position of the robot over time. Recently, Elinas and Little [7] presented a system that uses MCL in combination with a database of SIFT features learned in the same restricted environment. All these approaches use stereo vision to compute the 3D position of a landmark and match the visual features in the current view to all those in the database to ﬁnd correspondences. To avoid matching the observations to the whole database of features, we present a system that determines the sets of visible features for clusters of particles. These relevant features are then matched to the features in the current image. This way, the number of ambiguities, which can occur in larger environments, is reduced. The relevant features are determined by applying a ray-casting technique in the map of features. The main diﬀerence to existing metric localization systems using SIFT features is however that our approach is applicable to robots that are equipped with a single perspective camera only, whereas the other approaches require stereo vision or omnivision. Note that Davison et al. [3] and Lemaire et al. [13] presented approaches to feature-based SLAM using a single camera. These authors use extended

198

M. Bennewitz et al.

Kalman ﬁlters for state estimation. Both approaches have only been applied to robots moving within a relatively small operational range. Vision-based MCL was ﬁrst introduced by Dellaert et al. [4]. The authors constructed a global ceiling mosaic and use simple features extracted out of images obtained with a camera pointing to the ceiling for localization. Systems that apply vision-based MCL are also popular in the RoboCup domain. In this scenario, the robots use environment-speciﬁc objects as features [19, 22].

3 Monte-Carlo Localization To estimate the pose xt (position and orientation) of the robot at time t, we apply the well-known Monte-Carlo localization (MCL) technique [5], which is a variant of Markov localization. MCL recursively estimates the posterior about the robot’s pose: p(xt | z1:t , u0:t−1 ) = η · p(zt | xt ) ·

xt−1

p(xt | xt−1 , ut−1 ) · p(xt−1 | z1:t−1 , u0:t−2 ) dxt−1 (1)

Here, η is a normalization constant resulting from Bayes’ rule, u0:t−1 denotes the sequence of all motion commands executed by the robot up to time t − 1, and z0:t is the sequence of all observations. The term p(xt | xt−1 , ut−1 ) is called motion model and denotes the probability that the robot ends up in state xt given it executes the motion command ut−1 in state xt−1 . The observation model p(zt | xt ) denotes the likelihood of making the observation zt given the robot’s current pose is xt . To determine the observation likelihood, our approach compares SIFT features in the current view to those SIFT features in the map that are supposed to be visible (see Section 5). MCL uses a set of random samples to represent the belief of the robot (i) about its state at time t. Each sample consists of the state vector xt and a (i) weighting factor ωt that is proportional to the likelihood that the robot is in the corresponding state. The update of the belief, also called particle ﬁltering, is typically carried out as follows. First, the particle states are predicted according to the motion model. For each particle a new pose is drawn given the executed motion command since the previous update. In the second step, new individual importance weights are assigned to the particles. Particle i is (i) weighted according to the likelihood p(zt | xt ). Finally, a new particle set is created by resampling from the old set according to the particle weights. Each particle survives with a probability proportional to its importance weight. Due to spurious observations it is possible that good particles vanish because they have temporarily a low likelihood. Therefore, we follow the approach proposed by Doucet [6] that uses the so-called number of eﬀective particles [14] to decide when to perform a resampling step

Metric Localization with SIFT Features Using a Single Camera

Neﬀ =

1 N i=1

w(i)

2,

199

(2)

where N is the number of particles. Neﬀ estimates how well the current particle set represents the true posterior. Whenever Neﬀ is close to its maximum value N , the particle set is a good approximation of the true posterior. Its minimal value 1 is obtained in the situation in which a single particle has all the probability mass contained in its state. We do not resample in each iteration, instead, we only resample each time Neﬀ drops below a given threshold (here set to N2 ). In this way, the risk of replacing good particles is drastically reduced.

4 Acquiring 2D Maps of Scale-Invariant Features We use maps of visual landmarks for localization. To detect features, we use the Scale Invariant Feature Transform (SIFT). Each image feature is described by a vector p, s, r, f where p is the subpixel location, s is the scale, r is the orientation, and f is a descriptor vector, generated from local image gradients. The SIFT descriptor is invariant to image translation, scaling, and rotation and also partially invariant to illumination changes and aﬃne or 3D projection. Lowe presented results illustrating robust matching of SIFT descriptors under various image transformations [15]. Mikolajczyk and Schmid compared SIFT and other image descriptors and showed that SIFT yields the highest matching accuracy [17]. Ke and Sukthankar [10] presented an approach to compute a more compact representation for SIFT features, called PCA-SIFT. They apply principal components analysis (PCA) to determine the most distinctive components of the feature vector. As shown in their work, the PCA-based descriptor is more distinctive and more robust than the standard SIFT descriptor. We therefore use that representation in our current approach. As suggested by Ke and Sukthankar, we apply a 36 dimensional descriptor vector resulting from PCA. To acquire a 2D map of SIFT features, we used a B21r robot equipped with a perspective camera and a SICK laser range ﬁnder. We steered the robot through the environment to obtain image data as well as proximity and odometry measurements. The robot was moving with a speed of 40cm/s and collected images at a rate of 3Hz . To be able to compute the positions of features and to obtain ground truth data, we used an approach to grid-based SLAM with Rao-Blackwellized particle ﬁlters [8]. Using the information about the robot’s pose and extracted SIFT features out of the current camera image, we can estimate the positions of the features in the map. More speciﬁcally, we use the distance measurement of the laser beam that corresponds to the horizontal angle of the detected feature and the robot’s pose to calculate the 2D position of the feature. Thus, we assume that the features are located on the obstacles detected by the laser range ﬁnder. In the oﬃce environment in

200

M. Bennewitz et al.

which we performed our experiments, this assumption leads to quite robust estimates even if there certainly exist features that are not correctly mapped. In each 2D grid cell, we store the set of features that are supposed to be at that 2D grid position. Currently, we use a grid resolution of 10 by 10cm. In the ﬁrst stage of mapping, we store all observed features. After the robot moved through the environment, the number of observed SIFT features is extremely high. Typically, we have 150-500 features extracted per image with a resolution of 320 by 240 pixels. This results in around 600,000 observed features after the robot traveled for 212m in a typical oﬃce environment. After map acquisition, we down-sample a reduced set of features that is used for localization. For each grid cell, we randomly draw features. A drawn feature is rejected if there is already a similar feature within the cell. We determine similar features by comparing their PCA-SIFT vectors (see below). We sample a maximum of 20 features for each grid cell. Using the sampling process, features that were observed more often have a higher chance to be selected and features that were detected only once (due to failure observations or noise) are eliminated. The goal of this sampling process is to reduce computational resources and at the same time obtain a representative subset of features. To choose good representatives for the features, a clustering based on the descriptor vectors can be carried out. The left image of Figure 3 shows a 2D grid map of SIFT features of an oﬃce environment that was acquired by the described method. The ﬁnal map contains approximately 100,000 features. Note that also a stereo camera system, which was not available in our case, would be an appropriate solution for map building. The presented map acquisition approach is not restricted to robots equipped with a laser range ﬁnder.

5 Observation Model for SIFT Features In the previous section, we described how to built a map of SIFT features using a robot equipped with a camera and a proximity sensor. In this section, we describe how a robot without a proximity sensor can use this environmental model for localization with a single perspective camera. Sensor observations are used to compute the weight of each particle by estimating the likelihood of the observation given the pose of the particle in the map. Thus, we have to specify how to compute p(zt | xt ). In our case, an observation zt consists of the SIFT features in the current image: zt = {o1 , . . . , oM } where M is the number of features in the current image. To determine the likelihood of an observation given a pose in the map, we compare the observed features with the features in the map by computing the Euclidean distance of their PCA-SIFT descriptor vectors. In order to avoid comparing the features in the current image to the whole set of features contained in the map, we determine the potentially visible

Metric Localization with SIFT Features Using a Single Camera

201

features. This helps to cope with an environment that contains similar landmarks at diﬀerent locations (e.g. several similar rooms). In case one matches the current observation against the whole set of features, this leads to serious errors in the data association. To compute the relevant features, we group close-by particles to a cluster. We determine for each particle cluster the set of features that are potentially visible from these locations. This is done using ray-casting on the feature grid map. To speed-up the process of ﬁnding relevant features, one could also precompute for each grid cell the set of features that are visible. However, in our experiments, computing the similarity of the feature vectors took substantially longer than the ray-casting operations. Typically, we have 150-500 features per image. In order to compare two SIFT vectors, we use a distance function based on the Euclidian distance. The likelihood that the two PCA-SIFT vectors f and f belong to the same feature is computed as p(f = f ) = exp −

f −f 2 · σ12

,

(3)

where σ1 is the variance of the Gaussian. In general, one could use Eq. (3) to determine the most likely correspondence between an observed feature and the map features. However, since it is possible that diﬀerent landmarks exist that have a similar descriptor vector, we do not determine a unique corresponding map feature for each observed feature. In order to avoid misassignments, we instead consider all pairs of observed features and relevant map features. This set of pairs of features is denoted as C. For each pair of features in C we use Eq. (3) to compute the likelihood that the corresponding PCA-SIFT vectors belong to the same feature. (i) This information is than used to compute the likelihood p(zt | xt ) of an (i) observation given the pose xt of particle i, which is required for MCL. Since a single perspective camera does not provide depth information, we can use only the angular information to compute this likelihood. We therefore consider the diﬀerence between the horizontal angles of the currently observed (i) features in the image and the features in the map to compute p(zt | xt ). More speciﬁcally, we compute the distribution over the angular displacement of a particle given the observation and the map. For each particle, we compute a histogram over the angular diﬀerences between the observed features and the map features. The x-values in that histogram represent the angular displacement and the y-values its likelihood. The histogram is computed using the pairs of features in C evaluated using Eq. (3). In particular, we compute for each pair (o, l) ∈ C the diﬀerence between the horizontal angle at which the feature was observed and the angle at which the feature should be located according to the map and the particle pose. We add the likelihood that these features are equal, which is given by Eq. (3), to

202

M. Bennewitz et al.

the corresponding bin of the histogram. As a result, we obtain a distribution about the angular error of the particle. In mathematical terms, the value h(b) of a bin b (representing the interval of angular diﬀerences from α− (b) to α+ (b)) in the histogram is given by p(fo = fl ),

h(b) = β + (o,l)∈C

(4)

α− (b)≤α(o)−α(l)<α+ (b)

where α(·) is the function that computes the horizontal angle of a feature for a given pose of the robot, fo is the PCA-SIFT descriptor of feature o, and fl of feature l accordingly. β is a constant greater that zero ensuring that no angular displacement has zero probability. The histograms of particles that are close to the correct pose of the robot have high values around zero. In case that there are several similar features in the environment, the histogram has multiple modes. One ﬁnally needs to compute the observation likelihood of a particle. So far, we computed the distribution about the horizontal angular displacement, not its actual value. In case of a uni-modal or Gaussian distribution it would be suﬃcient to consider only the distance of the mean from zero taking into account the variance. However, in real-world situations, it is likely that one obtains multi-modal distributions. Each bin of that histogram stores the probability mass of the corresponding angular displacement of the particle. Therefore, we compute the observation likelihood given we have the angular displacement of that bin and multiply it with the value stored in that bin. The observation likelihood given the histogram is then computed by the sum over these values (i)

p(zt | xt ) =

h(b) · exp − b

α+ (b) + α− (b) 1 · 2 2 · σ2 2

2

,

(5)

where σ2 is the variance of a Gaussian describing the likelihood of a particle depending on the angular displacement. Figure 1 illustrates the whole process of computing the observation likelihood for a single particle. Note that a further improvement of the sensor model can be obtained by using the joint compatibility test between pairs of feature as proposed by Neira and Tard´ os [18] and not considering all possible data associations.

6 Experimental Results To evaluate our approach to estimate the pose of the robot equipped with a single perspective camera, we carried out a series of real-world experiments with wheeled and humanoid robots in an oﬃce environment. The B21r robot that performed the mapping task carries a standard camera with an opening angle of approximately 65◦ . In order to show that the acquired feature map

Metric Localization with SIFT Features Using a Single Camera 0.1

1 0.8

0.06

weight

likelihood

0.08

0.04 0.02 0

203

0.6 0.4 0.2

-3

-2

-1

0

1

2

3

angular displacement [rad]

(a)

0

-3

-2

-1

0

1

2

3

angular displacement [rad]

(b)

weighted likelihood

0.1 0.08 0.06

According to Eq. (5), (i) this leads to p(zt | xt ) = 0.25.

0.04 0.02 0

-3

-2 -1 0 1 2 3 angular displacement [rad]

(c) Fig. 1. Image (a) shows the distribution about the horizontal angular displacement for a particular particle computed according to Eq. (4). The plot shown in (b) depicts the Gaussian that is used to compute the weight of a sample depending on the displacement. Finally, image (c) shows the resulting histogram in which each bin of the histogram (a) is multiplied by the corresponding value of the Gaussian. Summing up the bins leads to an observation likelihood of 0.25.

can be used by robots equipped with diﬀerent cameras, we performed the localization experiments using a low-cost wide-angle camera (with an opening angle of about 130◦ ). The diﬀerence between typical images of both cameras can be seen in Figure 2. The arrows indicate the location, orientation, and scale of the generated SIFT features. The acquired map is depicted in Figure 3. 6.1 Localization Accuracy In this experiment, the wheeled robot traveled a distance of approximately 20m. Figure 3 shows the estimated trajectory as well as the true pose of the robot during this experiment. The ground truth has been determined using laser range data. The evolution of the particle ﬁlter is illustrated in Figure 4. It shows the particle clouds as well as the true position and the pose estimate provided by odometry.

204

M. Bennewitz et al.

Fig. 2. Example images with generated SIFT features. The images were obtained from two diﬀerent cameras used in the experiments. The standard camera (left) was used for map acquisition as well as for localization and a low-cost wide-angle camera (right) for further evaluation of our localization approach. features

weighted mean true pose odometry

5

0

y [m]

y [m]

5

-5 -10

0

-5 -10

-5

0 x [m]

5

10

-5

0

5

10

x [m]

Fig. 3. The left image shows the 2D map acquired in a typical oﬃce environment. Each cross represents the estimated 2D position of a SIFT feature. The right image depicts the estimated trajectory as well as the ground truth of a localization experiment. As can be seen, the weighted mean of the particles is close to the true pose of the robot.

A more quantitative analysis showing the localization error over time can be found in Figure 5. Between time step 40 and 50, the error in the pose of the vehicle was comparably high. This is because we used the weighted mean of the samples for the error computation and because the belief was temporarily multi-modal. This fact can be observed in the snapshots depicted in Figure 4. As this experiment illustrates, our technique is able to accurately estimate the pose of the robot. The average error in the x/y-position was 39cm. The average error in the orientation of the vehicle was 4.5◦ . We got comparable localization results when using diﬀerent cameras with a more constrained ﬁeld of view like the one which was used for map acquisition. During our

Metric Localization with SIFT Features Using a Single Camera true pose

true pose

odometry

205

true pose

odometry

t=0

t=12

t=33

true pose

true pose

true pose

odometry

odometry

odometry

t=41

odometry

t=60

t=50

Fig. 4. The particle set during localization. The two arrows indicate the pose resulting from odometry information as well as the true pose of the robot. The true pose of the vehicle was determined by using a laser range ﬁnder that was mounted on the robot for this purpose. The occupancy grid map is only shown for a better illustration and was not used for localization. odometry weighted mean

4 3 2 1 0

0

10 20 30 40 50 60 70 80

10 angular error [deg]

position error [m]

5

weighted mean

5 0 -5 -10

0

10 20 30 40 50 60 70 80

time

time

Fig. 5. Evolution of the error during the localization experiment depicted in Figure 3.

experiments, we used 800 particles in our particle ﬁlter, which were initialized with a Gaussian centered at the starting pose of the robot.

6.2 Tracking the Pose of a Humanoid Robot To further evaluate our approach, we applied our localization technique to the humanoid robot depicted in Figure 6. To estimate the pose of the robot

206

M. Bennewitz et al.

Fig. 6. The humanoid robot Max.

based on executed motion commands, we perform dead reckoning. The control input consists of the gait target vector that controls the lateral, sagittal, and the rotational speed of omnidirectional walking. The estimated velocities are integrated to determine the relative movement. Compared to a wheeled robot equipped with odometry sensors, this leads to a noisy pose estimate. Furthermore, due to the design of the humanoid robot, the camera images are often blurred because of vibrations.

plot 1

plot 2 plot 3

Fig. 7. The trajectory of Max. The red circles indicate the positions where observations were made. The corresponding plots of the particle clouds are shown in Figure 8.

In this experiment, the robot Max traveled along the trajectory shown in Figure 7. The red circles correspond to position where an observation was made. The particle clouds obtained in this experiment are given in Figure 8. In case no sensor information is integrated, the pose estimate has a high uncertainty as can be seen in the ﬁrst row of that ﬁgure. In contrast to this,

Metric Localization with SIFT Features Using a Single Camera plot 1

plot 2

207

plot 3

Fig. 8. Vision-based localization of a humanoid robot. The images in the ﬁrst row depict the evolution of the particles in case no sensor information is used. The high uncertainty in the particle cloud results from the poor motion estimate resulting from dead reckoning. The images in the second row show the result of our localization approach. As can be seen, the visual information allows to accurately estimate the pose of the humanoid robot.

the use of our vision-based localization technique reduces the uncertainty and enables to localize the humanoid. Note that due to unstable motion of the humanoid, missing odometry sensors, vibrations, and the shaking camera, the localization is less robust compared to a wheeled robot.

7 Conclusions In this paper, we presented an approach to mobile robot localization that relies on a single perspective camera. Our technique is based on Monte-Carlo localization and uses SIFT features extracted from camera images. In the observation model of our particle ﬁlter, we compare descriptor vectors of features in the current image to the set of potentially visible map features given the pose of the particles. Based on this information, we compute a distribution about the angular displacement for each sample given the current observa-

208

M. Bennewitz et al.

tion. The evaluation of potential correspondences between features is done eﬃciently by performing the necessary computations for clusters of particles. By using only the relevant features in the vicinity of the particles in the observation model, we reduce the number of data association failures. As we demonstrate in real-world experiments carried out with a wheeled as well as with a humanoid robot, our system provides an accurate metric pose estimate for a mobile robot without requiring proximity sensors, omnivision, or a stereo camera.

Acknowledgment This project is partially supported by the German Research Foundation (DFG), grant BE 2556/2-1 and SFB/TR-8 (A3). We would like to thank D. Lowe for providing his software to detect SIFT features and Y. Ke for his PCA-SIFT implementation. Further thanks to J. St¨ uckler and M. Schreiber for helping us carrying out the experiments with the humanoid robot.

References 1. H. Andreasson, A. Treptow, and T. Duckett. Localization for mobile robots using panoramic vision, local features and particle ﬁlter. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2005. 2. T.D. Barfoot. Online visual motion estimation using FastSLAM with SIFT features. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2005. 3. A.J. Davison, Y. Gonz´ alez Cid, and N. Kita. Real-time 3D SLAM with wideangle vision. In IFAC/EURON Symposium on Intelligent Autonomous Vehicles (IAV), 2004. 4. F. Dellaert, W. Burgard, D. Fox, and S. Thrun. Using the Condensation algorithm for robust, vision-based mobile robot localization. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 1999. 5. F. Dellaert, D. Fox, W. Burgard, and S. Thrun. Monte Carlo localization for mobile robots. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 1998. 6. A Doucet. On sequential simulation-based methods for bayesian ﬁltering. Technical report, Signal Processing Group, Departement of Engeneering, University of Cambridge, 1998. 7. P. Elinas and J.J. Little. σMCL: Monte-Carlo localization for mobile robots with stereo vision. In Proc. of Robotics: Science and Systems (RSS), 2005. 8. G. Grisetti, C. Stachniss, and W. Burgard. Improving grid-based SLAM with Rao-Blackwellized particle ﬁlters by adaptive proposals and selective resampling. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2005. 9. H.-M. Gross, A. K¨ oning, C. Schr¨ oter, and H.-J. B¨ ohme. Omnivision-based probabilistic self-localization for a mobile shopping assistant continued. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2003.

Metric Localization with SIFT Features Using a Single Camera

209

10. Y. Ke and R. Sukthankar. PCA-SIFT: A more distinctive representation for local image descriptors. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2004. 11. J. K˘ os´ecka and L. Li. Vision based topological Markov localization. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2004. 12. L. Ledwich and S. Williams. Reduced SIFT features for image retrieval and indoor localization. In Australian Conf. on Robotics and Automation (ACRA), 2004. 13. T. Lemaire, S. Lacroix, and J. Sol` a. A practical 3D bearing-only SLAM algorithm. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2005. 14. J.S. Liu. Metropolized independent sampling with comparisons to rejection sampling and importance sampling. Statist. Comput., 6:113–119, 1996. 15. D. G. Lowe. Object recognition from local scale-invariant features. In Proc. of the Int. Conf. on Computer Vision (ICCV), 1999. 16. E. Menegatti, M. Zoccarato, E. Pagello, and H. Ishiguro. Image-based MonteCarlo localisation with omnidirectional images. Robotics & Autonomous Systems, 48(1):17–30, 2004. 17. K. Mikolajczk and C. Schmid. A performance evaluation of local descriptors. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2003. 18. J. Neira and J. D. Tard´ os. Data association in stochastic mapping using the joint compatibility test. IEEE Transactions on Robotics and Automation, 17(6):890– 897, 2001. 19. T. R¨ ofer and M. J¨ ungel. Vision-based fast and reactive Monte-Carlo Localization. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2003. 20. S. Se, D.G. Lowe, and J.J. Little. Global localization using distinctive visual features. In Proc. of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), 2002. 21. R. Sim, P. Elinas, M. Griﬃn, and J.J. Little. Vision-based SLAM using the RaoBlackwellised particle ﬁlter. In IJCAI Workshop on Reasoning with Uncertainty in Robotics (RUR), 2005. 22. M. Sridharan, G. Kuhlmann, and P. Stone. Practical vision-based Monte Carlo localization on a legged robot. In Proc. of the IEEE Int. Conf. on Robotics & Automation (ICRA), 2005. 23. J. Wolf, W. Burgard, and H. Burkhardt. Robust vision-based localization by combining an image retrieval system with Monte Carlo Localization. IEEE Transactions on Robotics and Automation, 21(2):208–216, 2005.