SuperMimic Reproduced on One GPU Beats UC Berkeley's H100 Result

A GitHub developer using the handle zita-ch has released supermimic_vibed, an unofficial reproduction of the SuperMimic humanoid-control method. The maintainer reports writing the initial implementation with an AI coding assistant in about an hour. A subsequent 1.47-hour training run on one RTX PRO 6000 reached a normalized evaluation return of 0.829 across 40 motion clips, with every clip clearing the repository’s success threshold.

The repository covers the controlled LAFAN1 experiment from Figure 4a of the paper, where SuperMimic reached 0.82 after 1.2 hours on an H100. The paper’s broader 129,514-motion benchmark remains outside the reproduction’s reported scope.

The maintainer calls the project a “speed-vibecoded reproduction” and says the code has not been reviewed. It is an independent, MIT-licensed release from a personal account rather than a lab organization. The one-hour coding claim documents prototyping speed only; review, repeated runs and independent validation remain separate work.

One world model, many motions

The SuperMimic paper, titled SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model, is by Liao, Shao, Long, Zakka, Horowitz, Abbeel and Sreenath at UC Berkeley. It addresses humanoid motion tracking, where a control policy learns to reproduce human motion-capture clips on a robot.

Common model-free reinforcement-learning pipelines either train specialists for individual motions and compress them through distillation, which can lose behavior, or train one general policy with substantial simulator throughput. Both approaches repeatedly gather on-policy experience from the physics simulator.

SuperMimic trains a shared, differentiable world model from simulator transitions. Because the motion clips use the same robot and rigid-body dynamics, one learned model can support many clips. The policy improves by backpropagating through short imagined trajectories generated by that model. The paper calls this process analytic policy gradients, or APG.