The Problem of the Modeled World: Why Embodiment Requires Internal Simulation

The Problem of the Modeled World: Why Embodiment Requires Internal Simulation

An agent moving through the physical world confronts a problem that does not arise for disembodied reasoners: every action it takes irrevocably alters the state it will next perceive, and it must commit to such actions under conditions of partial observability and noisy measurement. A purely reactive policy — one that maps observations directly to motor commands — can succeed in narrow, well-structured environments, but it cannot, in general, anticipate the consequences of its own interventions, integrate evidence over time, or reason about entities it has never directly observed. These capacities are precisely those that distinguish competent physical intelligence from competent pattern matching. The claim of this book is that they are inseparable from the construction and maintenance of an internal model — a learned, generative, predictive representation of environmental dynamics that we will call a *world model*. This opening chapter establishes why such a commitment is not a contingent architectural preference but a structural necessity for any embodied system that must act coherently across time.

We proceed in five movements. First, we examine the embodiment hypothesis in cognitive science and argue that it implicitly entails a commitment to internal simulation. Second, we trace the historical lineage of world modeling from Helmholtzian unconscious inference through cybernetic feedback to contemporary latent predictive architectures. Third, we formalize the distinction between model-free and model-based control and show why the latter dominates whenever the environment is partially observable, stochastic, or non-stationary. Fourth, we decompose any embodied agent into four functional modules — perception, representation, prediction, and control — and show that the world model inhabits the representational core. Finally, we present a worked example in which two robots, one reactive and one model-grounded, are subjected to identical navigation disturbances; the divergence in their trajectories demonstrates that internal prediction is not merely convenient but constitutive of the more sophisticated behavior.

1.1 The Embodiment Hypothesis and Its Unstated Commitment

The embodiment hypothesis, as it has developed since the late 1980s through the work of Varela, Thompson, and Rosch, and as it has been reformulated in robotics by Brooks, Pfeifer, and Bongard, holds that cognition is not a disembodied symbol manipulation process but is constituted by the sensorimotor coupling between an organism and its environment. The hypothesis appears, on its face, to be a thesis about the *insufficiency* of computational cognition in the classical sense: it warns against treating the body as an input/output periphery to a central reasoner. But the claim has a less discussed converse. If cognition is constituted by sensorimotor coupling, then the structure of the coupling — the way in which sensory signals are lawfully related to motor outputs through the dynamics of the body and the world — must itself be *known* to the agent in some usable form. A coupling whose consequences are unknown cannot be exploited; it can only be reacted to.

This observation is rarely stated explicitly in the embodiment literature, perhaps because it sounds suspiciously like a return to the very internalism the hypothesis was meant to displace. But the world-modeling framework that has emerged in the past decade — through the work of Ha and Schmidhuber on latent dynamics, of Finn and Levine on visual foresight, of Pathak and colleagues on curiosity-driven prediction, and of the broader model-based reinforcement learning community — does not contradict embodiment. It completes it. The world model is not a homunculus sitting inside the agent contemplating a symbolic replica of reality; it is the agent's *implicit competence* regarding the regularities of its own sensorimotor loop, encoded as a parametric function approximator and updated through interaction.

Three clarifications are necessary at the outset.

First, a world model is not a photograph of the world, nor is it a complete generative simulator of reality at the level of physics. Such a model would be computationally intractable and observationally inaccessible. A world model is a *task-relevant abstraction*: it captures those aspects of environmental dynamics that are predictive of the consequences of the agent's own actions, at the resolution required for the decisions it must make. A navigation system needs a model of free space, friction, and inertial dynamics; it does not need a model of leaf flutter or ambient temperature, however causally real these may be. The abstraction is learned, not given, and it is shaped jointly by the data the agent encounters, the inductive biases of its architecture, and the objective function under which it is optimized.

Second, the world model is *internal* in a specific technical sense: it is a function implemented by the agent's own computational substrate, mapping observations and candidate actions to predicted outcomes, and it is updated on the basis of the agent's own experience. This distinguishes it from a model supplied a priori by a designer, though such designer-supplied structure often forms the initial scaffolding on which learning operates. The agent does not have epistemic access to the true environment; it has access to the world only through the model, which is itself a fallible, compressed, and selective representation.

Third, the world model is *predictive* in the sense that its principal functional role is to forecast the consequences of action: given the current state and a candidate action, what will the next state look like, what reward will accrue, and how will the distribution of possible futures be shaped? Prediction is the operational lingua franca through which perception, representation, and control communicate. Perception extracts the information necessary to update the model's current state estimate; representation is the space in which predictions are made; control selects actions on the basis of those predictions.

These three properties — task-relevant abstraction, internality, and predictiveness — together constitute the working definition of a world model that will be used throughout this book. We will refine and formalize the definition in subsequent chapters, but for the moment it suffices to note that a system possessing these properties can, in principle, exhibit behaviors that a purely reactive system cannot: it can plan, it can imagine counterfactual scenarios, it can credit past actions with downstream consequences, and it can transfer knowledge from one task to another without re-learning the underlying dynamics.

The embodiment hypothesis, properly understood, *requires* such a model. The body is not a peripheral device; it is part of the dynamical system whose evolution the agent must anticipate in order to act coherently. To act coherently is to select actions whose predicted consequences are favorable. To predict consequences is to possess a model. The conclusion is structural, not empirical: any agent that acts coherently over time in a partially observable environment implements, implicitly or explicitly, a world model. The interesting question is not whether to have one, but how to represent, learn, evaluate, and use it well.

1.2 A Genealogy of Internal Modeling

The contemporary framework of world modeling did not emerge ex nihilo. It is the inheritor of a long intellectual tradition in which the problem of inferring hidden structure from observable signals, and of using that structure to predict and act, has been repeatedly reformulated as the technical apparatus of the time permitted. Three intellectual movements deserve particular attention: Helmholtzian perception as unconscious inference, the cybernetic theory of feedback and internal models, an