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

SUMMARY

Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel and Koushil Sreenath have submitted a paper called SuperMimic to arXiv, dated 20 October 2026. It has been accepted at CoRL 2026 as an oral presentation. The paper addresses the cost of training humanoid whole-body controllers that track large human-motion datasets.

The authors say current pipelines rely on model-free reinforcement learning. Some merge per-motion specialist policies through lossy distillation; others train a single policy with cluster-scale compute. SuperMimic instead learns one differentiable world model of the dynamics the motions share, fitted from logged transitions of a standard rigid-body simulator. The world model and the tracking policy are trained together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients.

In 3 hours on a single GPU, the authors say SuperMimic exceeds the return of a state-of-the-art motion-tracker baseline trained for 7 days on 128 GPUs, which they describe as a reduction of more than 1000 times in GPU-hours. On the benchmark of 129,514 motions it fails on 25, against 179 failures for the baseline. After 9.8 hours on two GPUs it reaches zero failures.

On scaling, the authors report that from 1 to 264 hours of motion the world model's prediction error stays stable, and the GPU-hours needed to reach a target return grow slowly. They also state that the policy transfers to a physical Unitree G1 across diverse and acrobatic motions.

Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.