Moravec’s paradox

On this page 7

Moravec’s paradox is the observation that high-level reasoning requires relatively little computation, while low-level sensorimotor skills such as perception, grasping and walking require enormous computational resources. Artificial intelligence researchers noticed in the 1980s that programs could play chess or prove theorems at an expert level, yet no machine could match the perception and mobility of a one-year-old child.

The idea is named after roboticist Hans Moravec, who stated it together with Rodney Brooks, Marvin Minsky and others. It remains one of the core ideas in AI fundamentals because it explains why automation progress is uneven: abstract cognitive work often yields to software sooner than physical, embodied work.

The evolutionary explanation behind Moravec’s paradox

Moravec explained the paradox through evolutionary time. Sensorimotor skills were refined over hundreds of millions of years of natural selection, so the brain performs them through vast, highly optimized circuitry that works below conscious awareness. Abstract reasoning, by contrast, is a recent evolutionary addition of perhaps a few hundred thousand years, and humans perform it slowly and with visible effort.

Because people feel effort only during deliberate thought, they misjudge task difficulty. Walking across a cluttered room feels trivial, yet it hides massive unconscious computation: depth estimation, balance control, contact prediction and continuous error correction. Long division feels hard, yet it reduces to a short algorithm that a computer executes in microseconds.

The table below contrasts how humans experience a task with how difficult the same task has proven for machines.

TaskEffort felt by humansDifficulty for machinesReason
Chess at master levelHighSolved by search in 1997Small, fully specified rule set
Symbolic integrationHighSolved by 1960s programsFormal rules and transformations
Recognizing a faceNegligibleUnreliable until the 2010sHigh-dimensional, variable input
Folding laundryLowStill largely unsolvedDeformable objects and uncertain contact
Walking on rough groundNegligibleSolved only in constrained settingsReal-time balance and terrain sensing
Perceived human effort versus historical machine difficulty

Computational evidence and Moravec’s estimates

Moravec supported the argument with rough compute estimates derived from the retina. He reasoned that the retina performs edge and motion detection equivalent to roughly 1,000 million instructions per second, and he extrapolated that matching the whole brain would require on the order of 100 million MIPS. These figures are order-of-magnitude estimates, not measurements, and later neuroscience has both supported and disputed the details.

The deeper point concerns where the complexity resides. Reasoning programs operate on clean symbols that a programmer has already extracted from the world. Perception must build those symbols from noisy, ambiguous sensor data, which is exactly the step that early symbolic AI quietly assumed away. This gap is closely related to the symbol grounding problem and the frame problem.

The following table summarizes the main lines of evidence usually cited for the paradox.

Evidence typeObservationWhat it suggests
Neural resourcesLarge cortical areas serve vision and motor controlPerception carries heavy computation
Early AI resultsLogic and game programs succeeded quicklyFormal reasoning is compact
Robotics historyManipulation stalled for decadesPhysical interaction is costly
Development in childrenGrasping and walking precede abstract reasoningSensorimotor skill is foundational
Evidence commonly cited for Moravec’s paradox

How the paradox holds up in modern AI

Modern deep learning has reshaped the picture but has not erased it. Convolutional networks made image classification reliable once large labeled datasets and graphics processors became available, which shows that perception was not impossible but data-hungry and compute-hungry. Large language models now handle language and code fluently, sometimes before robots can reliably pick up an unfamiliar object.

The paradox therefore persists most clearly in embodied tasks. Robots must act in real time, cope with friction, deformable materials and occlusion, and learn from expensive physical data that cannot be scraped from the web at scale. Simulation helps, yet the sim-to-real gap often degrades performance when a learned policy meets real hardware.

This table shows how the status of typical tasks has shifted over the decades.

Task categorySymbolic era statusDeep learning era status
Formal reasoningStrongStrong, now also probabilistic
Visual recognitionWeakStrong with large datasets
Natural languageBrittleFluent, with factual errors
Dexterous manipulationVery weakImproving, still narrow
Open-world mobilityVery weakReliable only in mapped environments
Changing status of tasks across eras of AI research

Moravec’s paradox is often discussed together with other foundational ideas, but each makes a distinct claim. Confusing them weakens analysis of where AI will advance next.

ConceptCore claimRelation to the paradox
Moravec’s paradoxSensorimotor skill is computationally expensiveBaseline idea
Bitter lessonGeneral methods with scale beat hand-crafted knowledgeExplains how perception was finally cracked
Commonsense reasoningEveryday knowledge is hard to encodeA cognitive cousin of the same difficulty
Symbol grounding problemSymbols need links to perceptionNames the missing step in symbolic systems
AI-completeSome problems need general intelligenceLabels tasks that resist narrow solutions
Moravec’s paradox compared with neighboring ideas

Practical implications and common misreadings

For practitioners, the paradox offers a useful heuristic for automation forecasting. Tasks performed on screens with structured inputs tend to be automated earlier, whereas tasks requiring fine motor control in unstructured environments tend to resist automation longest. Plumbing, caregiving and cleaning illustrate this pattern better than accounting or translation.

  • Misreading one: the paradox does not claim that reasoning is easy in general; it claims that formalized reasoning is cheap once the problem is already symbolic.
  • Misreading two: it is not a mathematical law but an empirical generalization, and exceptions exist in both directions.
  • Misreading three: recent progress in perception does not refute it, because that progress required enormous data and compute, which is the paradox’s own prediction.
  • Misreading four: language fluency does not imply physical competence, so benchmarks on text say little about robotic capability.

The simple explanation

Computers find the things people consider hard surprisingly easy, and the things people consider easy surprisingly hard. A computer can beat a world champion at chess, but it struggles to fold a towel or walk across a messy room.

The reason is that humans have practiced seeing, moving and grabbing for an extremely long evolutionary time, so the brain does it automatically and people never notice the effort. Chess is new, so it feels hard. In short, what feels effortless to people often hides the most complex work.

How it is used in practice

A logistics company planning automation in a mid-sized warehouse applies the paradox when ranking projects. Software for inventory forecasting and route planning is deployed within a few months, because those tasks run on structured data in existing systems.

Robotic picking of mixed items from bins is treated as a separate, longer program. The team runs a pilot with a vision-guided arm on about 2,000 product types and finds that rigid boxes are picked reliably, while soft bags and transparent packaging cause frequent failures. The company keeps human pickers for those items and plans the budget around a gradual rollout. The outcome is a realistic automation roadmap that avoids overpromising on physical manipulation.