Behavior-Based Robotics By Ronald C. Arkin Cambridge, MA, and London, UK: MIT Press, 1998 491 pages. ISBN 0-262-01165-4 Reviewed By Henry Hexmoor Department of Computer Science University of North Dakota Grand Forks, ND 58202 hexmoor@cs.und.edu Mobile robotics programming has been a field without a textbook. Ronald Arkin has written a book that fairly exhaustively compiles research related to mobile robotics. The book title ``Behavior-Based Robotics'' is named after the most popular paradigm in mobile robotics and the book content embodies the most important work in the paradigm. Inspired by Ethology, Psychology, and Neurology, behavior-based robotics paradigm engineers a robot's tasks into units of behavior and organizations of such behaviors called architecture. In all, this well written book is a superb addition to the field and will serve as an introduction and a reference guide. This book is well-suited as a semester long introductory course in mobile robotics programming at the senior level or the first year graduate level. The book is divided into 10 chapters. Chapters have a set of clearly laid our objectives and summary. There are no homework or execrises and the book does not come with a instructor's guide. The first three chapters discuss the origins of behaviors and tenets of behavior-based paradigm. Chapter 1, "Whence Behavior?", reviews the emergence of behavior-based robotics in literature. Arkin cites a definition for a "robot" but does not offer a definition for a mobile robot. His definition does include mobile robots but it is too inclusive of non-mobile robots. Theer are times that we want to know what is a mobile robot. Such a defintion would point to mobility by the entire robot, programmability of the robot, and use of sensors and perception for nontrivial reactions. Arkin reviews Braitenberg vehicles and the early mobile robot, Shakey. Chapter 2, "Animal Behavior", presents animal behaviors as models for building computational data structures for robot behavior. Arkin examines origins of behavior in Neurology, Psychology, and Ethology. Arkin's reference to "motivated behaviors" and "innate releasing mechanisms" were fascinating and sent me looking for references. Chapter 3, "Robot Behavior", does a great job of synthesizing a model that integrates various behavioral arbitration schemes. Arkin describes various methods of behavior arbitration, from subsumption style to summation. It is surprising that the word Animat is left out and there is only a casual reference to the field of ALife later in the book. Arkin tries to include only embodied robotic systems in is book and that leaves out most of ALife systems since they are simulated. Chapter 4, "Behavior-Based Architectures", contains a catalog of various architectures with pros and cons for each. Although knowledge engineering would show suitability of an architecture to a task, researchers tend to be dogmatic about their architecture. Arkin remains objective and balanced in his presentation and instead of synthesizing a unified architecture, he presents an unbiased evaluation of architectures. In addition to strengths and weaknesses, he introduces a set of neutral charateristics for each architecture. For instance, "robustness", and "timeliness of development" are neutral characteristics for the subsumption architecture. Chapter 5, "Representational Issues for Behavioral systems", is on knowledge representational issues. Arkin starts by reviewing knowledge representation and then turns to the need to model long and short term memory. The start of behavior based robotics coincided with reactive approach to robotics. However, unlike purely reactive robotics, behavior-based robots clearly benefit from internally modeling the world as we can see from representations of maps and in visual tasks. Chapter 6, "Hybrid Delberative/Reactive Architectures", covers a recent trend in building hybrid systems. Parts of the system are built in the tradition of Good Old Fashioned AI (GOFAI) while other parts are fashioned after reactive systems. GOFAI gives us abilities like planning and reasoning that are generally slow but accurate. Reactive approach provides behaviors that are generally fast and inexact. The hybrid system provides best of both worlds. Chapter 7, "Perceptual Basis for Bahavior-Based Control", presents a new paradigm of Perception that has evolved mostly from Computer Vision and is compatible with the behavior-based robotics. This new paradigm advocates distributed and need-to-know decomposition of perceptual tasks. Behaviors can be thought of as having a motor and perceptual component. Arkin reviews a few robotic sensors and outlines computer vision research supporting this new paradigm of perception. Chapter 8, "Adaptive Bahavior", is on machine learning. Arkin teaches us that habituation is a form of adaptation, and adaptation is a form of learning. He discusses learning issues with emphasis and examples from reinforcement, neural net, genetic algorithm, and fuzzy logic approaches. It is difficult to cover these learning techniques in one chapter but Arkin manages fairly well. Whereas reinforcement learning is a type of learning, Arkin's categories of neural net, genetic, and fuzzy control are representational schemes and not learning schemes. Also, Arkin ends up putting some learning techniques such as case-based learning, memory-based-learning, and explanation-based learning under other types of learning. We consider these to be unsupervised learning techniques. A way of categories learning schemes is: (1) reinforcement (same as Arkin's), (2) supervised (some neural net schemes), and (3) unsupervised (some forms in neural net, genetic, and fuzzy control schemes). Chapter 9, "Social Behavior", covers multi-robotics. Arkin begins by superbly motivating the need and niche for multiple robots working in the same environment. This chapter is fun to read as the reader can think of examples not included such as additions to Arkins Table 9.1 on the modes of animal communication. Some AI such as Speech Acts is not included in his discussion on communication. This chapter reviews many interesting projects such as Lynn Parker's Alliance architecture. We find his conclusions about communication content to be too terse and contain unexplained terms such as "display behavior". It probably seemed obvious to Arkin since this a major focus of his ongoing research. Chapter 10, "Fringe Robotics: Beyond Behavior", is on issues in th emargins of robotics. It includes some classical philosophical discussions from AI such as the Turing Test, Consciousness, and Searle's Chinese Room argument. This can be a source of endless discussions and essays. Arkin does a good job of including an introduction for such issues. As he puts it, these are not related to behavior-based robotics and are "beyond behavior". For a good collection of online papers in this area see David Chalmers' page http://ling.ucsc.edu/~chalmers/. Also, [Haugeland, 1997] is a good reference. Arkin includes other discussions on fringe of robotics such as emotions, Homeostatis, and Nanotechnology. These too are not behavior-based robotics. Arkin's book is a much needed textbook in mobile robotics. For those interested in hands on mobile robotics, Jones and Flynn's "Mobile Robots" book is a great book. Arkin's book is a great companion to "Mobile Robots" that expands Jones and Flynn's Chapter 9 on robot programming to a book-length presentation. Although Arkin intends to include manipulators and other industrial robots most of the presentation is on mobile robots. Clearly, behavior-based robotics can and should be applied to robots that not strictly mobile. Finally, In this reviewer's opinion, the book needs a section on Control Theory and how the AI approach to robotics can benefit from rigors of Control Theory. The interested reader may examine [Dean and Wellman, 1991]. If you seek to learn more about sensors, see Everett's "Sensors for Mobile Robots". Athough Arkin talks about sensors but not in enough depth to susbtitute [Everett, 1995]. We have all seen many of the robot pictures and famous work presented in this book on various television programs. A video collection of carefully selected segments will be a nice companion to this book. In fact, this might be a useful endeavor as a future activity. References T. Dean and M. Wellman, 1001, Planning and Control, Morgan Kaufman Publishers, CA. J. Jones and A. Flynn, 1993. Mobile Robots: Inspiration to Implementation, AK Peters, Ltd, MA. H.R. Everett, 1995. Sensors for Mobile Robots, AK Peters, Ltd, MA. J. Haugeland, 1997. Mind Design II: Philosophy, Psychology, and Artificial Intelligence, MIT Press, MA.