The Embodiment Hypothesis

In 1990, roboticist Rodney Brooks published a paper that would reshape how researchers think about intelligence. He argued that the classical AI project — build a rich internal model of the world in a computer, reason over it, then output instructions to a body — had the relationship exactly backwards. Intelligence, he claimed, is not something that happens inside a box and then gets expressed through a body. Intelligence emerges from the interaction between a body and the world. The body is not the vehicle for a mind; the body is part of the mind. This claim, refined and formalized over the following decades, became known as the embodiment hypothesis.

The Classical AI Assumption and Its Critics

Classical AI — sometimes called Good Old-Fashioned AI, or GOFAI — assumed a clean separation between a reasoning system and its body. The reasoning system maintains a symbolic world model: a map of objects, their locations, their properties, and the rules governing how actions change them. Given a goal, the system plans a sequence of actions by searching this symbolic model for a path from the current state to the goal state. The body simply executes the plan. Brooks's critique was brutal and empirical. He built a series of insect-like robots — creatures named Allen, Herbert, and Genghis — that had no central world model at all. Each behavior (avoid obstacles, walk, track a moving object) was implemented as a separate reactive layer that read sensors and produced motor commands directly, with no intermediate symbolic representation. These robots moved through real cluttered environments with a robustness that GOFAI systems of the same era completely lacked. The lesson Brooks drew: for real-world physical interaction, the world is its own best model. Use sensors to read what is there rather than maintaining an error-prone internal copy. This was a provocation, not a final answer. But it seeded a research tradition — embodied, situated, behavior-based robotics — that permanently changed the field.

Cognitive Science Roots: Merleau-Ponty and Enactivism

Brooks was a roboticist, but the philosophical groundwork for the embodiment hypothesis was laid decades earlier in cognitive science and phenomenology. French philosopher Maurice Merleau-Ponty argued in the 1940s that perception is not a passive reception of data — it is an active, bodily engagement with the world. When you reach for a glass, your hand does not first compute a trajectory and then execute it. The act of reaching is itself a form of knowing. Your body has pre-reflective skills, what Merleau-Ponty called motor intentionality, that do not pass through explicit reasoning. Cognitive scientists Francisco Varela, Evan Thompson, and Eleanor Rosch developed this into a framework called enactivism: the idea that cognition is not computation inside a brain, but action in the world, structured by the continuous coupling between organism and environment. An organism does not process a pre-existing world — it enacts a world through its patterns of sensorimotor activity. For AI researchers, enactivism has a concrete implication: if cognition is constitutively shaped by the sensorimotor history of a particular body in a particular environment, then an AI system without a body cannot fully replicate human cognition — it can only approximate the outputs of that cognition through statistical patterns in recorded human behavior.

Implications for Modern AI

The embodiment hypothesis does not say that disembodied AI is useless — clearly it is not. Large language models and image classifiers accomplish remarkable things. But the hypothesis makes specific predictions about what such systems will struggle with. First, common-sense physical reasoning. A language model trained on text learns that 'a glass balanced on a table edge will fall if pushed.' But it learns this as a linguistic pattern, not through any history of knocking objects off tables. When novel physical scenarios arise that are not covered by the training distribution, disembodied systems often fail in ways that a toddler — who has embodied experience of gravity and balance — would not. Second, dexterous manipulation. Picking up a fragile egg requires integrating visual information about shape and position with tactile feedback about contact force and slip detection, all in real time. This is a task that even the most sophisticated large models cannot directly perform — they have no hands, no tactile sensors, no motor system to train. Third, adaptive physical navigation. Walking over rough terrain, squeezing through a narrow gap, or catching a thrown object requires real-time integration of proprioception (knowledge of your own body's position and motion), vision, and balance — capabilities that emerge from a history of physical interaction, not from reading about it. The embodiment hypothesis, if correct, means that truly general intelligence will require embodied systems — robots that learn not just from data, but from acting in the physical world.