Why Bodies Are Hard

Moravec's paradox

In his 1988 book Mind Children, the roboticist Hans Moravec observed that it is comparatively easy to make computers perform like adults on intelligence tests or at checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. Marvin Minsky summed up the same insight as easy things are hard. Writing in 2021, the computer scientist Melanie Mitchell listed the opposite assumption, that easy things are easy and hard things are hard, as one of the common fallacies in predicting AI progress.

Intelligence is not only in the brain

Mitchell names another fallacy: that intelligence is all in the brain. Human intelligence, she argues, is a strongly integrated system, bound up with the body, emotions and common sense. A 2022 survey of embodied AI in IEEE Transactions on Emerging Topics in Computational Intelligence describes a shift in research from internet AI, learned from data online, to embodied AI, where agents learn through interacting with their surroundings from a first-person view. It rests on the belief that true intelligence can emerge from that interaction.

Brooks and the case for building up

The roboticist Rodney Brooks pushed this view in his 1991 paper Intelligence without Representation. He argued that a robot need not hold an explicit model of the world or of its own intentions, and that its behaviours need no central controller. His subsumption architecture stacked simple reactive layers; in his robot Allen, each layer acted only when the layer below was idle. His earlier slogan that the world is its own best model comes from a 1990 paper. In a 2007 assessment, the philosopher Vincent Müller judged the defining feature of Brooks's proposals to be their architecture.

What it means for superintelligent robots

Even if reasoning in software races ahead, the hard parts of robotics are where Moravec pointed: seeing clearly, gripping safely, keeping balance and recovering from small mistakes. Simulation helps, but the 2022 survey covers simulated agents, and moving skills from simulation to real machines remains its own challenge. For anyone choosing robots, the lesson is practical. A high score on an AI benchmark says little about how a robot will handle your objects, floors and lighting, so test the physical work itself.

Sources and further reading

• Why AI is Harder Than We Think — Melanie Mitchell, Santa Fe Institute (arXiv).
  Four fallacies in forecasting AI, quoting Moravec's paradox in full and making the case for embodied cognition.
• Is there a Future for AI without Representation? — Minds and Machines, Vincent C. Müller (arXiv).
  Peer-reviewed assessment of Rodney Brooks's 1991 paper Intelligence without Representation and its subsumption design.
• A Survey of Embodied AI: From Simulators to Research Tasks — IEEE Transactions on Emerging Topics in Computational Intelligence (A*STAR).
  Survey from Singapore's A*STAR on the shift from internet AI to embodied AI and the simulators used to train agents.

Common questions

What is Moravec's paradox in simple terms?

Skills that feel hard to people, like chess or exams, have proved easier to automate than skills a toddler has, like seeing, grasping and moving around a room.

Does Moravec's paradox still hold?

Researchers still cite it. Robot learning is improving, but everyday physical skills remain hard to automate reliably, so progress on text and images is not a good guide to progress in robots.