Skip to content

SuperMimic: Single-GPU Motion Learning Edition

Choose a tag to compare

@kozzy kozzy released this 24 Oct 10:47
· 1 commit to master since this release

Train a G1 motion tracker from scratch on one GPU. This special edition integrates SuperMimic into HoloMotion, bringing learned world models and differentiable imagined rollouts to the existing motion-data and training workflow.

Our selected MuJoCo Warp recipe achieved 80% success on 40 complete LAFAN1 training clips after 30 minutes on a single RTX 5090.

Highlights

  • Learn through a differentiable world model. Fit dynamics from physical transitions, then optimize the policy through multi-step imagined rollouts.
  • Train on one CUDA GPU. The validated compact actor has approximately 9 million parameters.
  • Choose your simulator. Run SuperMimic with MuJoCo Warp or Isaac PhysX.
  • Compare against PPO. Shared task and evaluation interfaces support comparisons using the same motion inputs, observations, rewards, and policy architecture.
  • Reuse the HoloMotion workflow. Work with HDF5 motion data, G1 assets, Hydra configurations, and the existing training entrypoint.
  • Inspect the evidence. Recorded configurations, training curves, checkpoints, input fingerprints, and physical evaluation transitions accompany the workflow.

Recorded Results

Single RTX 5090 · 30-minute training budget · seed 0 · 40 complete LAFAN1 training clips.

Method MuJoCo Warp Isaac PhysX
SuperMimic 32/40 (80%) 24/40 (60%)
Reference PPO 0/40 (0%) 0/40 (0%)

Success means completing a clip without triggering the height or tilt termination rule. These are training-set evaluations; success does not impose a global tracking-error threshold.

Getting Started

Notes

Based on tag supermimic, commit adaacd50.

The results compare a selected SuperMimic recipe with reference PPO configurations, rather than independently tuned best configurations for both methods. They are single-seed measurements and do not establish held-out generalization, real-robot transfer, or an algorithm-wide advantage.

MuJoCo supports training recovery. Isaac supports fresh training and final evaluation; checkpoint recovery and periodic evaluation are currently unavailable.

Acknowledgements

Credit for the method belongs to the SuperMimic paper. This integration adapts supermimic_vibed, with its MIT notice retained, and uses SuperMimic-Reproduced as an independent cross-check.