Foundations: From Symbolic AI to Embodied Cognition

Foundations: From Symbolic AI to Embodied Cognition

An agent that does not possess an internal theory of how its actions reshape the world is forced to rediscover that theory every time it acts. Classical artificial intelligence, captured in the slogan "the intelligent behavior is the rational manipulation of symbolic representations," deliberately abstracted away from this concern: it inherited from formal logic the assumption that perception delivered ready-made propositions and that action consisted in selecting operators whose preconditions were satisfied and whose effects were already known. Six decades of empirical evidence — from Shakey's brittleness in cluttered hallways to the asymptotic sample complexity of deep reinforcement learning in sparse-reward environments — show that this abstraction is the precise source of the field's long-standing inability to build agents that operate robustly in the physical world. This chapter reconstructs the intellectual trajectory that produced the modern "world model" paradigm and establishes, in formal terms, why such an internal model is a computational necessity rather than a stylistic preference.

1.1 The Symbol Grounding Problem and the Necessity of Embodiment

Harnad's 1990 critique, *The Symbol Grounding Problem*, put a precise name on an intuition that practitioners had carried for two decades: a symbol, in the formal sense inherited from Carnap, Tarski, and Newell, is a token whose meaning is fixed entirely by its relational role within a closed symbolic system. There is nothing in the syntax of `BLOCK(x) ∧ ON(x, y) ∧ CLEAR(y) → MOVE(x, y)` that ties `BLOCK` to the physical decay of wooden cubes or `MOVE` to the mechanical reach of a manipulator. The mapping is supplied, by hand, by the human designer. The formal symbol `MOVE` does not *stand for* the act of lifting; it merely *co-occurs* with the human's mental token for lifting. Carnap's *explicated* notion of meaning was an explicit admission that the grounding must be performed by an *explicator* external to the formalism — typically, the engineer.

The temptation to dismiss the critique on the grounds that some sufficiently comprehensive symbolic system must "bottom out" in sensor readings is precisely what Harnad warns against. In any closed symbolic system — and the universal Turing machine argument guarantees that every symbolic AI is, at bottom, a closed symbolic system — the intrinsic meaning of the symbols is zero. The meaning is parasitic on the implementation. Embodied agents escape this paradox precisely because their representational states are transduced: at the boundary between sensor and silicon, there is a physical process whose dynamics are not exhausted by the symbolic system. Pixel intensities, microphone voltages, joint torques, and bumper switches are not "intrinsically" meaningful either, but they are *constrained* by the physics of the world in a way that the symbols of a planner are not. Transduction is not grounding, but it is the necessary substrate for grounding.

Two consequences follow and shape the rest of the book.

First, robustness. The reason a STRIPS planner fails the moment a sensor returns `BLOCK(x)` with a probability distribution over `x` rather than a definite value is that its entire inferential engine assumes the symbolic input is correct. Replace the discrete sensing assumption with the actual conditional observation density $p(o_t \mid s_t)$, and the planning problem becomes intractable in the worst case. World models, by contrast, are trained against the empirical observation density, and they degrade gracefully: the agent remains functional when perception is noisy, incomplete, or adversarially perturbed.

Second, sample efficiency. A pure model-free learner, when dropped into a new environment, must repeatedly execute the consequence of every action to learn its expected effect. A model-based learner can compress one execution of a transition into a learned conditional $p(s_{t+1} \mid s_t, a_t)$, from which it can then simulate arbitrary counterfactuals. In environments where physical interaction is expensive (robotics, autonomous driving, protein design) or dangerous (clinical decision-making, space exploration), this distinction is not academic; it is the difference between a usable system and an unbuildable one.

The embodiment thesis that animates this book is therefore not the naive claim that "agents should have bodies," but the precise claim that the learning of intelligent behavior requires a closed loop between a sensorimotor interface and a learnable internal dynamics model. The remainder of this chapter builds the formal apparatus that makes this claim precise, traces its historical lineage, and motivates the modern learned-world-model paradigm as its most recent and most empirically successful instantiation.

1.2 A Brief Intellectual Archaeology: From STRIPS to Learned World Models

The history of embodied AI is, in a real sense, a sequence of failed attempts to postpone the grounding problem. Each generation of researchers conceded a little more of the closed-world assumption, replaced a little more of hand-engineered structure with learned components, and discovered that the concession was non-trivially costly. The path from STRIPS to DreamerV2 is not linear — there are genuine dead ends, including the early 1990s abjection of neural networks — but every subsequent advance is intelligible as a response to a limitation that its predecessor exposed.

1.2.1 Classical AI and the STRIPS Era (1966–1985)

Shakey the robot, designed at SRI between 1966 and 1972, is the canonical first embodied agent in the modern sense. Its perception stack, the vision module Hearsay-II descendant HWM (Harris-Hunt-Winkler), produced symbolic scene descriptions; its planner, STRIPS (Stanford Research Institute Problem Solver), searched for sequences of operator applications that transformed an initial state into a goal state; its execution module, PLANEX, compiled a plan into run-time scripts that called low-level behavioral routines. The STRIPS assumption — that each operator $a \in \mathcal{A}$ is described by an *add list*, a *delete list*, and a precondition list, all expressed as conjunctions of positive literals — encoded an extraordinarily strong inductive bias: the world is closed, finite, deterministic, fully observable, and discretizable. Shakey worked precisely to the extent that the hallway at SRI honored those assumptions, and failed precisely to the extent that the world refused to be discretized (cables, shadows, uneven door thresholds, other people).

The intellectual achievement of STRIPS was not the planner itself but the *assumption of completeness*. By assuming that the state description was grounded, that actions had known effects, and that perception delivered precise updates to the world model, the planner reduced embodied intelligence to search in a discrete graph. The General Problem Solver lineage (Newell, Shaw, Simon, 1959) had already shown that search could, in principle, automate reasoning; STRIPS demonstrated that it could, in principle, automate action. The assumption of completeness was a Faustian bargain. It bought tractability at the price of brittleness, and the brittleness became intolerable as soon as researchers tried to leave the lab.

1.2.2 The Subsumption Revolution (1985–1995)

Rodney Brooks' subsumption architecture was a wholesale reaction to STRIPS. Brooks' diagnosis was unfriendly but accurate: intelligence, as a capacity for symbolic reasoning, had been confused with intelligence as a capacity for situated behavior. A creature that hunts successfully in a tidal flat operates without a STRIPS database and without a state estimator; it operates through a layered set of competence modules, each of which is a finite-state machine over tightly scoped sensor channels, and which collectively *subsume* lower layers rather than calling them