Event

SuperMimic: Scaling Humanoid Whole-Body Control with a Single Differentiable World Model

1 reports · 1 sources

Qiayuan Liao et al. propose SuperMimic, which trains a differentiable world model and a tracking policy simultaneously in a single training loop, providing analytic gradients to the policy via backpropagation-through-time over short imagined trajectories. The method amortizes learning across all motions with a single differentiable world model of shared dynamics, fitted from transition data logged by a standard rigid-body simulator. The paper claims that SuperMimic, trained for 3 hours on a single GPU, exceeds the return of an advanced motion-tracking baseline (which was trained for 7 days on 128 GPUs); after 9.8 hours of training on two GPUs it reaches zero failures. The policy has been transferred to a physical Unitree G1 robot, covering diverse and acrobatic motions. The paper was accepted as an Oral at CoRL 2026.