Physical Intelligence

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI confronts a technology that has mastered the symbolic and remained helpless in the material. Language models can describe how to tie a knot and cannot tie one. They can write a flawless paragraph about catching a ball and cannot catch one. Rus’s insistence on physical intelligence is, among other things, a corrective to the temptation to mistake fluency for understanding and prediction for competence. The machine that passes every symbolic test may have no idea how to act in the world those symbols describe.

The concept also reframes the question of human distinctiveness in the AI era. The fluency that has most impressed and unsettled the public—writing, coding, reasoning over text—is precisely what the machines do best. What they do worst is what every living organism does as a matter of course: navigate physical space, handle objects, adapt to surprise. This asymmetry suggests that the distinctively human contribution, in a world where AI handles the symbolic, may be more embodied, more physical, more relational than our cognitive self-image has recognized.

Origin

Rus has used the term and championed the concept throughout her career at MIT’s CSAIL, where she has pursued soft robots, self-reconfiguring machines, and liquid neural networks as converging approaches to the embodied intelligence problem. Her framing has become more urgent as the dominant strategy in AI has tilted decisively toward ever-larger language models—systems that improve rapidly at the symbolic and do relatively little to close the gap in physical competence.

The concept also has deep roots in robotics research more broadly. The insight that physical intelligence requires a different kind of architecture from symbolic or statistical intelligence goes back at least to Rodney Brooks’s behavior-based robotics in the 1980s and 1990s, and to the phenomenological tradition in philosophy that argued intelligence is fundamentally embodied rather than representational. Rus’s contribution has been to make this claim concrete, in working machines, at a level of ambition and experimental rigor that previous generations could not achieve.

Key Ideas

The Asymmetry of Easy and Hard. The tasks that feel easiest to humans are hardest for machines—grasping an irregular object, walking across uneven terrain, making sense of a cluttered room—while the tasks that feel hardest to humans, like doing calculus or playing chess, were the first to fall. This “Moravec’s paradox” reveals something deep about the nature of physical intelligence: it is older, more deeply evolved, more distributed through the body, and more reliant on real-time feedback from a physical world that does not cooperate.

Embodied Intelligence Distributes Through the Body. In embodied systems, not all of the intelligence has to be in the processor. A soft gripper that conforms to the shape of what it grasps does not need to compute the object’s geometry; its material properties absorb that complexity. An octopus distributes control into its arms. A bird navigates turbulence through aerodynamics its nervous system did not explicitly model. Physical intelligence can be morphological—built into the shape and material of the body—as well as computational.

Biological Inspiration as a Research Strategy. Nature has solved physical intelligence at scales of efficiency and robustness that artificial systems have not approached. A fly navigates a room with a brain smaller than a grain of rice. Caenorhabditis elegans finds food and avoids harm with 302 neurons. These are not curiosities but clues: proof-of-concept demonstrations that embodied intelligence can be achieved with radical compactness, and that the path to robust physical AI may run through understanding what biology has already discovered.

The Cloud-vs.-Body Question. Physical intelligence cannot depend on the cloud. A robot navigating a disaster site, a car deciding whether to brake, a device operating in a remote environment: all require intelligence that lives inside the machine itself, running on modest hardware, without network access. This is one of the motivations behind Rus’s liquid neural networks and her broader argument for compact, embedded models—the conviction that intelligence designed to act in the physical world must be carried by the system that acts, not borrowed from a distant facility.