A thought about robot hands

A baby spends months learning to grab things.

Reaching, missing, reaching again. Fingers closing on air. Knocking a block sideways instead of picking it up. Slowly, over weeks and weeks, the brain and the hand learn to talk to each other. Eventually: a grip. The simplest thing. Fingers closing around an object with just enough force to hold it, not so much force that it breaks.

We spend months on this and then we never think about it again.

OpenAI’s Dactyl project taught a robotic hand to manipulate a Rubik’s cube. The hand can rotate the cube, flip it, move individual faces. It’s impressive. The dexterity is real.

But that same hand can’t pick up an egg. Not reliably. The force calibration for something fragile, something with an irregular surface, something that requires the kind of unconscious delicacy a toddler masters by age two, is still unsolved.

We’re amazing at what robots find hard. Robots are amazing at what we find easy.

I don’t have a big point here. I just think about this sometimes. The things we take for granted, the things so simple we don’t even register them as skills, are the things that keep the smartest engineers in the world up at night. And the things that impress us about machines (speed, precision, tirelessness) are trivial for them.

We’re mirrors of each other, in a way. Each remarkable at what the other can’t do.

I find that beautiful and confusing in equal measure.

Related thinking: