The Embodied Turn in Artificial Intelligence

The Embodied Turn in Artificial Intelligence

The question that animates this opening chapter cuts to the heart of a fundamental debate within artificial intelligence: why must intelligent systems be embodied, and what essential problems does the embodiment perspective solve that purely cognitive or computational approaches cannot? This question is not merely academic—it determines the architectural choices, research priorities, and ultimate aspirations of the field. For decades, mainstream AI pursued intelligence primarily through the manipulation of symbolic representations divorced from any physical substrate or sensory connection to the world. The limitations of this approach have become increasingly apparent, prompting a shift toward embodied paradigms that situate artificial agents within environments where perception and action are inextricably intertwined.

The embodiment turn represents more than a technical adjustment; it embodies a reconceptualization of what intelligence fundamentally is and how it emerges. Rather than viewing cognition as the abstract manipulation of symbols according to formal rules, embodied AI posits that intelligent behavior arises from the dynamic interaction between agent and environment—a process in which sensorimotor experience, physical constraints, and situated action play constitutive roles. Understanding this shift is essential for appreciating why world models have emerged as a central research priority and how they differ fundamentally from earlier representational approaches.

This chapter traces the intellectual and methodological trajectory from classical symbolic AI to contemporary embodied cognition, examining the theoretical and empirical motivations for this transformation. We begin with the historical foundations of symbolic AI, exploring both its achievements and its fundamental limitations. We then introduce the perception-action cycle as a foundational organizational principle for embodied systems, followed by an examination of sensorimotor contingencies and their role in shaping representations. The free energy principle and predictive processing frameworks provide a unifying theoretical architecture that connects these concepts and motivates the world modeling approach. Throughout, we ground these theoretical discussions in concrete examples and empirical findings, building toward an integrated framework that will guide the reader through the subsequent chapters of this book.

Historical Foundations: From Physical Symbol Systems to Embodied Cognition

The Rise and Dominance of Symbolic AI

The historical origins of artificial intelligence are deeply intertwined with the physical symbol systems hypothesis, articulated most influentially by Allen Newell and Herbert Simon at the 1956 Dartmouth conference that gave the field its name. According to this hypothesis, a physical symbol system—defined as a set of elementary physical tokens that can be combined to form larger structures and operated upon by processes that create, copy, modify, and destroy these structures—has the necessary and sufficient means for general intelligent action. This definition deliberately echoes the Turing machine model: symbols that can be manipulated according to rules are sufficient to generate any computable behavior, including the full range of human cognition.

The implications of this hypothesis were profound for how AI research would be conducted. If intelligence is fundamentally symbol manipulation, then physical instantiation is irrelevant—the same symbolic program should produce the same behavior whether running on a biological brain, a digital computer, or any other physical substrate. This abstraction from embodiment enabled researchers to focus exclusively on computational processes, treating the body as a mere peripheral input/output device and the environment as a passive backdrop against which cognitive operations unfold. The research program that emerged from this framework became known as "Good Old-Fashioned Artificial Intelligence" (GOFAI), and it dominated the field for several decades.

The achievements of symbolic AI were genuinely remarkable. The general problem solver (GPS), developed by Newell, Shaw, and Simon in the late 1950s, demonstrated that problem-solving behavior could be decomposed into means-ends analysis operating over symbolic representations. The Logic Theorist proved mathematical theorems by searching through proof spaces represented as symbolic structures. Expert systems of the 1970s and 1980s captured domain-specific knowledge in rule-based formats, achieving notable successes in areas such as medical diagnosis, chemical analysis, and geological exploration. The规划 systems (planning systems) developed by researchers such as Fikes and Nilsson showed how complex sequences of actions could be synthesized by reasoning backward from goal states through symbolic representations of the world.

Yet even during this period of apparent triumph, fundamental limitations were becoming apparent to those who looked carefully. The brittleness of symbolic systems—their tendency to fail catastrophically when confronted with situations that deviated even slightly from their explicit assumptions—posed a persistent challenge. The famous SHRDLU system, developed by Terry Winograd in the early 1970s, provided an impressive demonstration of natural language understanding within a constrained blocks world, but this performance depended entirely on the careful手工 construction of a world model perfectly aligned with the system's representations. When SHRDLU's assumptions were violated or its vocabulary extended beyond the designed scope, performance degraded rapidly. This brittleness was not an accidental feature that improved engineering might eliminate; it reflected a fundamental architectural choice that prioritized logical consistency over adaptive robustness.

The Symbol Grounding Problem

One of the most incisive critiques of the symbolic approach came from philosopher Stevan Harnad in 1990, who articulated what he termed the symbol grounding problem. Harnad observed that the semantic content of symbols in a computational system is parasitic upon the interpretations assigned by external users—there is no intrinsic connection between the arbitrary tokens manipulated by a symbol system and the meanings those tokens are meant to represent. A computational system can manipulate the symbol "water" in sophisticated ways, but unless the meaning of "water" is grounded in something beyond other symbols, the system does not truly understand what water is.

Consider the classic "Turing Test" scenario: a system that passes the test by producing text responses indistinguishable from a human might still have no genuine understanding of the concepts it discusses. The system manipulates symbols that correlate with meanings for us, but the correlations themselves do not constitute meaning for the system. This problem becomes particularly acute when we consider how a purely symbolic system could ever acquire new meanings or extend its understanding to novel domains. If all symbols derive their meanings from chains of reference back to other symbols, and if this chain must be established by external interpretation, the system can never autonomously bootstrap its own semantic content.

The symbol grounding problem highlights a deeper issue with disembodied cognition: without some non-symbolic interface with the world, symbols float free of the realities they purport to represent. The Chinese Room argument, articulated by philosopher John Searle in 1980, dramatized this point by imagining a person who, following formal rules, manipulates Chinese characters without understanding Chinese. Searle argued that formal symbol manipulation, however sophisticated, does not constitute understanding—the syntax is not sufficient for semantics. Whether one accepts Searle's specific conclusions or not, his argument drew att