SuperMimic cuts control compute by > 1,000x

A UC Berkeley paper, SuperMimic (accepted at CoRL 2026 and posted on arXiv on 20 October), shows a humanoid robot's whole-body controller can be trained on one GPU in three hours, versus roughly 21,000 GPU-hours for the leading published system SONIC - a more than 1,000x cut in control compute. Instead of trial and error in a physics simulator, SuperMimic first learns one model of how the robot's own body moves and then trains the controller inside that differentiable model: practice runs over 10x faster than the simulator at typical batch sizes (up to 62x at the largest), the differentiable model cuts practice steps needed by about 64x, and prediction error stays stable as motion data grows from 1 to 264 hours. On a benchmark of 129,514 motions it failed 25 times versus 179 for SONIC, and reached zero failures after 9.8 hours on two GPUs;