Concept

Athletic Intelligence

conceptroboticsai
Teach me this →

Athletic Intelligence

Athletic intelligence is Marc Raibert‘s term for the physical half of what makes a capable robot or animal: mechanical design, real-time motor control, and energetics working together to move, balance, and manipulate the world, as distinct from cognitive intelligence — the planning, perception, and judgement that decide what to do. Raibert coined the split to explain the mission of the Boston Dynamics AI Institute, where he argues Boston Dynamics had already, through decades of work on BigDog, Spot, and Atlas, set the standard for athletic intelligence, leaving cognitive intelligence as robotics’ real bottleneck.

Mechanism: what athletic intelligence is made of

Athletic intelligence is not raw hardware — Raibert is explicit that ‘people who think you don’t need to innovate hardware anymore are wrong’ — but hardware fused with a particular style of control. The clearest example is dynamic balance: a legged robot stays upright not by keeping enough feet planted to be stable at every instant (a table or chair’s stability), but by continuously predicting where its own motion is heading a second or two ahead and correcting for it in real time — the same trick a runner or a pole-vaulter uses. Raibert traces the mechanism back to his own pogo-stick robots of the early 1980s, where balance reduced to three coupled calculations: how much energy to add on each bounce, where to plant the foot relative to the body’s centre of mass, and how much corrective torque to apply to the body’s attitude while the foot is on the ground. Athletic intelligence, on this account, is the accumulated engineering skill of solving that kind of real-time physical control problem — for walking, running, manipulating objects, even dancing — reliably and across varied terrain.

The contrast with cognitive intelligence

Raibert’s dividing line matters because it cuts against a common assumption that a physically impressive robot is thereby a generally intelligent one. He offers a concrete example of the gap: on the way to an interview he consulted a map, estimated a walking time, and set his own departure time accordingly — ‘simple intelligence’, in his words, but exactly the kind of planning-under-everyday-uncertainty that most robots still cannot do for themselves. Most robots, he argues, remain ‘pretty dumb’ on this cognitive side, which is why they need extensive human programming for every task rather than being able to watch a person perform a task once and reproduce it — the ambition behind the AI Institute’s flagship project, Watch, Understand, Do. The practical consequence is that athletic capability alone does not make a robot commercially useful: reliability, cost, and especially cognitive competence, not leg mechanics, are what stand between a robot like Spot and broad deployment.

Where mainstream views differ

The more common framing in AI treats ‘intelligence’ as largely a cognitive or linguistic property — the capability large language models are judged against — with embodiment and control treated as a separate, lower-status engineering problem to be solved afterwards, or solved by scaling the same learned methods that work for language and vision. Raibert’s split pushes back on that hierarchy from the other direction: he regards engineering as, if anything, the senior discipline (‘scientists only get to study what’s out there, and engineers get to make stuff that didn’t exist before’), and treats athletic intelligence as a hard-won, already-mature achievement rather than a solved formality. He is also candid that the current best results in athletic control — Atlas’s dynamic manoeuvres, as of the interview — still lean on ‘traditional’ techniques such as model-predictive control rather than end-to-end learning, in some tension with the wider AI field’s expectation that learned methods will eventually subsume classical control everywhere. [?]

In the wiki