The disciplines of kinematics and dynamics of robots, spatial planning, connectionism, neuroscience, developmental psychology, and motor learning contributed to this work. It is a good example of what interdisciplinary approaches to AI can achieve, from both a research and a practical perspective. The book is an imaginative and worthwhile experience for anyone in any of the above research disciplines who feels intellectual myopia setting in and wants to step back momentarily and glimpse an approach that attempts to fuse diverse ideas.
The work deals with a connectionist system (Mel’s Ph.D. work) that learns the fundamental activity of reaching for a target object in a two-dimensional visual world with obstacles. Much to his credit, the author has implemented many of the components in a concrete fashion, rather than only in simulation. The system, nicknamed MURPHY, consists of a real-time binary vision system, a simple three-degrees-of-freedom robot arm that moves in the plane, and a workstation that simulates a large network of simple computing units. Mel makes a number of simplifying assumptions in his system. These are engineering compromises so that he can concentrate on the real issues at hand, namely “learning by doing” and mental imagery.
The book consists of eight chapters. The first chapter briefly defines the problem and gives a motivation for the study of visually guided reaching. A quick overview of existing research results that motivate the system design, and the assumptions that are made, follows.
Chapter 2 describes the vision hardware and electromechanical setup in more detail. The MURPHY system’s computational architecture and sensory modality representations are then discussed. The architecture consists of various functionally distinct subnetworks of simple computing units. In each grouping, the member units represent the state variables of its sensory modality. Subnetworks represent the visual field; the velocity of the hand; and the absolute angles and angular velocities of the shoulder, elbow, and wrist robot joints. Units have overlapping response profiles for a given scalar value, known to neurobiologists as coarse coding. This permits multiple noisy units to represent a continuous value with better precision than a single unit could. Mel also discusses MURPHY’s approach to inverse kinematics, or the computation of the desired joint angles and joint velocities for Cartesian positions and velocities.
Chapter 3 discusses how MURPHY learns. Mel uses the paradigm of “learning by doing,” in which the system flails its arm around in different configurations using joint-level commands and simultaneously observes the corresponding visual field activities. The system learns what effect joint angle perturbations will have on the visual field in the absence of visual stimuli. The development of this predictive mechanism (a form of mental imagery) is the core contribution of the approach, since associative learning between the joint angles and visual expectation makes concrete the abstract and tenuous notion of mental imagery. The nonlinear associational unit and its associated Hebbian learning algorithm are discussed, as are the relationships of this type of unit to higher-order neural receptive fields and lookup-table function learning paradigms.
Chapter 4 deals with the heuristics the system uses in avoiding obstacles around the target. The system uses three modes of planning, with increasing sophistication. All modes have innate visual routines that can check for collisions between arm links and obstacles in the environment. The first mode uses the joint-space to visual-space map imagery to simulate the effects of motions in its visual field representation. Random perturbations in the joint space are projected through the joint to visual space maps, and a perturbation that decreases the distance to the target in the associated image is sent to the actuators. This technique is therefore subject to being caught in local minima. The second, more sophisticated mode uses a depth-first backtracking search to escape local minima, and then sends the search output to the actuators. While the trajectories generated by MURPHY are initially suboptimal, looking somewhat like Brownian motion, they accomplish the job of getting to the target when possible. Most notably, when a third mode is used, performance improves rapidly by use of the associative mapping from Cartesian velocities in the visual field to joint velocities. To end the chapter, some issues relating to scaling the system to work in a full three-dimensional world are discussed.
Chapter 5 discusses how this research relates to existing solutions for robot spatial planning problems. Mel makes the point that even though his technique is not optimal, the computational burden of his search is much less than other techniques, and the method seems to be adequate in generating feasible trajectories to the target.
Chapters 6 and 7 discuss how the system can help to understand certain results from the vast literature concerning visually guided reaching from an experimental psychological and neurobiological perspective and how, in turn, this body of research helped Mel to formulate the design of the MURPHY system. Since the MURPHY architecture is essentially an embodiment of representational and computational mechanisms known to neuroscientists, such as coarse coding, receptive fields, Hebbian learning, higher-order receptive fields, and bidirectional connections between separate cortical areas, it is not surprising that many qualitative similarities exist between this system and previous results. Nevertheless, the relationships described are insightful and the analysis is thorough and comprehensive. Chapters 6 and 7 provide an excellent introduction to these biologically oriented disciplines for the engineer or computer scientist who is interested in how systems neuroscientists and experimental psychologists view representations and computations.
Chapter 8 concludes by discussing some of the practical lessons learned in the implementation, especially concerning empirical learning and the “learn only what you need” approach, the use of mental simulations, and scaling properties for storage and sample size in lookup table learning.
In general, Mel’s writing style is clear and easy to follow. He gives a good explanation of the psychological results he cites, although the chapter on biological ramifications requires frequent references to the map of cortical areas provided for those uninitiated in the Byzantine field of neuroanatomy. The sections discussing system scalability are somewhat speculative, and the author admits more work is needed in the analysis. Except for a couple of typographical errors, the book is well produced and illustrated, with a comprehensive reference section and index.
This work is a readable presentation of an exciting new approach to multisensory integration and learning for robots. It should interest researchers in a variety of disciplines.
