The best robotics video ever made involves a hockey stick

Boston Dynamics posted a video of the new Atlas this week, and I’ve watched it maybe fifteen times. It walks out of a building on its own and across snow, picks up boxes, and then a guy with a hockey stick keeps knocking the box out of its hands. It turns, finds the box again, and picks it back up. At the end someone shoves it from behind and it falls flat on its face, then gets up by itself.

A lot of the comments are about feeling sorry for the robot, which is funny and also a little revealing. It moves enough like a living thing that people project onto it.

What impresses me is the recovery, much more than the walking. Walking on flat ground can be done with a fairly rigid plan: you calculate where the feet go and move the body so its center of mass stays over the support area. That’s how most earlier humanoids walked, carefully, like they were on ice. Recovering from an unexpected push is a different problem. The robot has to sense that something went wrong, estimate how its whole body is moving and where the push is taking it, and choose a new step that catches the fall, all in a fraction of a second. On snow it also doesn’t know exactly how its foot will slip.

In mechanical terms, it has to keep running a model of its own dynamics against what its sensors report, and correct continuously when the two disagree. That’s balance. People do it so automatically that we don’t think of it as intelligence, but it’s probably one of the oldest and hardest skills a brain has. A toddler spends about a year getting it right.

The other thing worth noticing is that the new Atlas is smaller and lighter than the version in the DARPA challenge, and runs on its own battery with no tether. The DARPA robots last year fell over constantly on footage that went viral for the wrong reasons. This one looks like it came from a different decade.

I don’t know how much of this is learned versus hand-designed control. Boston Dynamics doesn’t publish much. My guess is it’s mostly carefully engineered control, not machine learning. Even so, it shows what the physical side of intelligence looks like when it’s done well. A robot that can shrug off a hockey stick understands the physics of its own body better than any vision system understands a photo.