Artificial intelligence (AI) is returning to its roots. Even before computers entered the public consciousness, science fiction promoted the image of robots with human-like intelligence, and issues such as sensory interpretation were a staple of early work in AI. Gradually, most researchers adopted an architectural perspective that posited clean boundaries between sensory acquisition, information processing, and actuators that change the state of the world. Most AI researchers focused their attention on the information-processing function, assuming that once perfected, their intelligent algorithms could easily interface with sensors and actuators to function in the real world.

This classic view has not been without dissenters. A decade ago, Rodney Brooks’s classic paper “Elephants Don’t Play Chess” argued that any useful concept of intelligence must be grounded in an agent’s relationship to the physical world. This perspective, known as embodied cognitive science, has produced an impressive array of results. Pfeifer and Scheier offer an integrated review of these results and a guide to designing situated intelligences. The book’s 19 chapters form six parts that lead the reader from classical concepts, through examples and design principles for embodied agents, to an agenda for future research.

Part 1 explores the foundations of the study of intelligence. It summarizes classical AI and cognitive science, then highlights a number of problems that frustrate the classical approaches, including the frame problem, the symbol grounding problem, and the challenge of real-world  interaction. 

Part 2 introduces basic concepts for the alternative approach offered by embodied cognitive science. It argues that intelligence must be studied using complete autonomous agents that integrate information processing with sensors and actuators. Central to the book’s approach is the “frame of reference problem,” which contrasts an agent’s internal mechanisms with the concepts attributed to the agent by an observer. For example, the high-level ontology used by an observer to describe an agent’s behavior may have no direct representation in the low-level specification of the agent itself, but may emerge from the interaction between the agent and its environment. Neural networks of various kinds are a core technology for constructing such agents, and this part includes a chapter introducing this technology.

Part 3 reviews different approaches to embodied cognition. It introduces the Braitenberg vehicles, an early thought experiment showing how high-level behaviors can emerge from the interaction of a low-level specification and the environment. Next, it discusses Brooks’s subsumption architecture, and then turns to approaches based on artificial life and artificial evolution. A final chapter briefly summarizes three other approaches: dynamical systems, behavioral economics, and schema-based mechanisms. Newcomers to the field will find this section of the book particularly helpful for the pointers that it offers to previous work in the field.

A major objective of the book is presenting design principles for autonomous agents. Part 4 outlines eight principles and expounds seven of them. The overarching “three-constituents principle” (discussed in detail in Part 5) asserts that design of an autonomous agent must include definition of its ecological niche and of the desired behaviors and tasks, as well as of the agent itself. Seven principles governing the morphology, architecture, and mechanisms of agents are discussed in this part. First, design should focus on complete agents, avoiding the classical AI approach of factoring out the “intelligent” part of the agent and leaving the interfaces to others. Second, many parallel, loosely coupled processes should interact asynchronously with an agent’s sensors and actuators. Third, all intelligence, including perception, categorization, and memory, should be conceived as sensory-motor coordination that structures sensory input. Fourth, designs should be kept cheap by exploiting the ecological niche. Fifth, interaction among redundant sensory channels can advance intelligent behavior. Sixth, the principle of “ecological balance” seeks to match the complexity of the agent with that of the task environment. Finally, the agent must have a value system that can drive self-supervised learning. This part concludes with a case study of human memory from the perspective of these principles.

Part 5 discusses how these principles interact in the overall design of an agent, thus developing the three-constituents principle, and provides guidelines for conducting and interpreting experiments with such agents.

Part 6 outlines future directions for embodied cognitive science. It discusses a series of hard problems for the next generation of researchers, revisits the definition of intelligence from an embodied perspective, and outlines the social implications of this approach to constructing intelligent systems.

The book is organized as a textbook. It emphasizes concepts rather than detailed algorithms, and so gives no programming or implementation exercises, but each chapter includes a list of thought-provoking issues that might serve as the basis for class discussions or student essays, a summary of key points, and a brief list of annotated references for further reading. An integrated glossary of over 150 terms, a comprehensive bibliography with over 350 entries through 1998, and separate author and subject indices help readers navigate the book’s encyclopedic bulk.

People trained in classical AI will find this book an articulate and thought-provoking challenge to much that they have taken for granted. People new to cognitive science will find it a stimulating introduction to one of the field’s most productive controversies. Pfeifer and Scheier deserve our thanks for a thorough, accessible, and courteous contribution in the best tradition of scholarly debate.