Reproduce “$\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation” 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

Weyl Lu, Yubei Chen

∇Mimic trains humanoid motion tracking from simulator gradients. It clips extreme per-sample gradients, converts them into action targets, and uses a GPU engine to speed gradient computation. On 40 LAFAN1 clips, it reaches normalized return 0.82 in about 20 minutes and 0.874 in 1.8 hours. The 0.82 comparison uses SuperMimic’s reported result from different hardware; experiments are limited to G1 simulation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.