Reproduce “SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel, Koushil Sreenath
SuperMimic uses a shared differentiable world model to scale humanoid whole-body motion tracking. It fits dynamics from simulator transitions, then improves one controller through imagined rollouts and analytic gradients, without differentiating the simulator. On a 129,514-motion benchmark, it exceeds SONIC’s 0.90 normalized return within 3 hours on one H100 GPU; the result is limited to training-motion coverage.
Key insights
- A single learned model can capture dynamics shared across a large motion library: prediction error stayed in a narrow range as the dataset grew from one to 264 hours of motion.
- Both faster imagined rollouts and gradient-based policy updates contribute to efficiency; in an ablation, the method matched PPO’s best training return with about 64 times fewer imagined steps.
- Model-based gradients optimize a learned surrogate rather than the simulator’s true dynamics, so model accuracy and transfer beyond tested settings remain important constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.