I'd Rather Have a Dog That Opens Doors Than a Humanoid Robot That Falls Over

Last week, while running offline tests on a scraping model in the lab, a colleague pulled me over to watch a video: Boston Dynamics' Atlas doing a backflip, then falling over, leaking hydraulic oil everywhere. Nearby, another quadruped robot dog was steadily climbing stairs, extending a robotic arm from its back to gently turn a door handle. Staring at that robotic arm's motion trajectory, I suddenly felt that what we quants talk about daily as "overfitting" applies equally to robotics—humanoid robots are those high-variance models that hit 99% accuracy on the training set but crash and burn on the test set.