What it is
From-scratch SuperMimic reproduction for humanoid motion tracking with a differentiable world model and BPTT policy training on mjlab/MuJoCo Warp. Train and evaluate Unitree G1 motion-tracking policies with a reproduced SuperMimic pipeline: data conversion, differentiable FK, learned world model, BPTT actor-critic, benchmarking, and replay.
Key results
- Hits 0.824 LAFAN1 score in 1.66 h on 1× RTX 5090
- World-model 8-step open-loop error is 0.0065-0.0099
- Runs ~2×10^5 env-steps/s at 4096 envs; 4-GPU efficiency is 91%
Shows SuperMimic-style differentiable world-model control can be reproduced at paper scale on commodity RTX 5090 hardware.