Moravec's paradox

Moravec's paradox names a stubborn mismatch between human and machine capability: tasks we find trivial, like grasping a coffee cup or walking across gravel, demand vastly more computational power from a robot than tasks we consider intellectually difficult, like playing chess or solving differential equations. A robot arm needs millions of sensor updates per second, complex feedback loops, and real-time collision detection just to pick up an object without crushing it. Meanwhile, teaching a machine to win at chess required brute-force calculation, yes, but nothing approaching the neural resources a toddler spends learning to walk.

The paradox arises from evolutionary mismatch. Human brains evolved over millions of years to solve sensorimotor problems first: survival meant seeing prey, grasping tools, moving through terrain. Abstract reasoning is comparatively recent, built on top of those embodied skills. We have specialized neural hardware for vision and motor control, honed by natural selection. Digital computers have no such inheritance. They excel at formal symbol manipulation, at following rules explicitly, but they lack the dense, parallel, analog substrate that makes human sensing and moving so efficient.

Where the bottleneck sits

Real-time sensorimotor control requires sampling rate, latency, and bandwidth that scale poorly. A robotic hand with 20 joints, 50 touch sensors, and camera input must compute joint torques, predict slip, detect collisions, and update strategy within 10 to 100 milliseconds. That is not one calculation. It is thousands of coupled differential equations, state estimation, and learned heuristics running in parallel. A chess engine, by contrast, can batch-process moves offline. Its latency can stretch to seconds. The difference is not subtle: it is the difference between closed-loop physical interaction and open-loop symbol shuffling.

The paradox has reshaped robotics research. Instead of building robots that reason about the world and then act, engineers now train neural networks on vision and proprioception, learning sensorimotor policies directly from data. This works better than explicit programming because the network can implicitly capture the statistics of the physical problem. Yet this shift, from logic to learning, still does not eliminate the core cost. A robot learning to grasp requires thousands or millions of trials, either in simulation or on hardware. A human infant grasps naturally within months.

Named after Hans Moravec in the 1980s, the observation remains central to debates about artificial general intelligence. It suggests that the hard problem of AI is not reasoning but embodiment. A machine that can hold a conversation but cannot open a door is not intelligent; it is a parlor trick with shallow roots in the real world. This has pushed robotics toward embodied learning, sim-to-real transfer, and biomimetic designs that borrow structure from biology rather than building from first principles.