Count the contacts

Count your own contact events

Pick an ordinary two-minute task: making a sandwich, unloading half a dishwasher, changing a duvet cover, tidying a drawer of cables.

Do it slowly and count every discrete contact event — each time a hand makes or breaks contact with an object. Count regrasps separately. Count the moments where you brace something with one hand while the other works.

Also note, roughly, how many of those contacts required force you were controlling rather than position — squeezing, not just holding.

Contact events in two minutes

Your count. Then divide by 120 for a per-second rate — that rate is what a general-purpose manipulator has to match to do your task at your speed.

Why legs came first

Locomotion is a control problem with a low-dimensional goal — stay upright, go that way — and a body whose dynamics you can model well. It is genuinely hard, and it is the kind of hard that responds to better control and better simulation, both of which arrived.

Manipulation is a contact problem. Contact is discontinuous, friction is poorly modelled, objects deform, and the thing you are holding occludes the thing you are looking at. Simulation transfers far less well, so the data has to come from the real world, slowly, through teleoperation and demonstration.

That asymmetry — not funding, not attention — is most of why you have seen a lot of very good walking and comparatively little very good folding.

Explain the asymmetry

Explain to somebody who has seen the backflip videos why folding laundry is the harder problem.