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

Authors: Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel, Koushil Sreenath · CoRL 2026

Licences

• Open Access
• Paper: arXiv.org perpetual, non-exclusive license 1.0

Under arXiv's licence, everything below is written from the abstract only.

Why it matters

Tracking large human-motion datasets is the main route to general humanoid whole-body control, but scaling it has meant lossy distillation of per-motion specialists or cluster-scale compute. The authors report surpassing a state-of-the-art tracker trained for 7 days on 128 GPUs after 3 hours on one GPU, with more than 1000x fewer GPU-hours and 25 failures on 129,514 benchmark motions against the baseline's 179.

Proposed solution

SuperMimic fits one differentiable world model of all motions' shared dynamics from logged rigid-body simulator transitions, and trains it together with the tracking policy in one loop, using backpropagation through time over short imagined rollouts to give the policy analytic gradients.

Abstract

Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than 1000× and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.

Links

• arXiv 2610.28480: https://arxiv.org/abs/2610.28480
• PDF: https://arxiv.org/pdf/2610.28480
• Code: https://github.com/HybridRobotics/supermimic
• Project page: https://supermimic.github.io/