robotics//embodied AI//Moravec's paradox

Moravec's paradox is the observation, made by the roboticist Hans Moravec and others in the 1980s, that what humans find intellectually hard (algebra, chess, logical proof) takes computers relatively little computation, while what any toddler or dog does without thinking (seeing, grasping, walking over rubble) takes an enormous amount, and it is used to explain why AI conquered board games and exams before it could fold laundry or empty a dishwasher. Moravec put it as computers reaching adult performance on intelligence tests while struggling to match a one-year-old in perception and mobility.

Moravec's paradox is the observation, made by the roboticist Hans Moravec and others in the 1980s, that what humans find intellectually hard (algebra, chess, logical proof) takes computers relatively little computation, while what any toddler or dog does without thinking (seeing, grasping, walking over rubble) takes an enormous amount, and it is used to explain why AI conquered board games and exams before it could fold laundry or empty a dishwasher. Moravec put it as computers reaching adult performance on intelligence tests while struggling to match a one-year-old in perception and mobility.

The usual explanation is evolutionary. Sensorimotor skill was refined over hundreds of millions of years and runs below awareness, so it feels effortless; abstract reasoning is recent, slow and conscious, so it feels hard, while the problem it solves is actually small and well specified. A chess position has a few dozen legal moves and exact rules. A hand closing on a glass deals with slippery surfaces, uncertain weight, noisy touch and a few milliseconds to react, all at once and with no rulebook.

An autopilot can hold a drone steady in gusts better than any pilot, and the same drone cannot pick up a cup from a cluttered table.

The first is a well-modelled problem in a few variables; the second is perception and contact in an open world.

In industry it decides where automation is easy. Scheduling a production line or optimizing a recipe is software; picking mixed parts out of a bin, handling soft packaging or plugging in a connector remains the costly cell that still needs a person or a very constrained setup.

It is an observation, and it is eroding. Learned controllers trained in simulation now walk quadrupeds over terrain hand-designed controllers struggled with, and manipulation is improving; what remains true is that each such gain needed data from interaction, which text cannot supply (embodied AI).

It is one face of uneven capability. A model that writes passable chemistry and misreads a simple physical scene is the same unevenness at a finer grain, described as the jagged frontier.