SuperMimic independent pretrained policy replay — 2026-11-12
Independent implementation: NJU-R-L-Group-Embodied-Lab/SuperMimic-Repro-Sim2Real
Revision f8f60d7bbdba68468c050ef100af02d752ca238e (MIT, Yuxi Liu / CuboiLeo and Hongyu Ding).
Motion Data by Bones Studio — https://bones.studio
BONES-SEED license: https://bones.studio/info/seed-license
No model weights or reference trajectories are redistributed here.
Use the upstream releases v1.0 (e13i_0291000) and v1.1 (e24_0355000_lean35kfs), subject to applicable licenses.
Python 3.10.12; mujoco 3.17.0; numpy 2.2.6; onnxruntime 1.23.2; scipy 1.15.3; pandas 2.3.3; pillow 12.3.0.
Unitree model revision 1eb6642e3f3fdfb7fb13a9794fd6a2dd93ea0e7d.
The wrapper calls the upstream G2 rollout, capturing actual free-base physics states. Rendering replays recorded states; it does not modify the evaluated physics.
Run python capture_policy.py --help and python render_policy.py --help for arguments. Set G1_UNITREE_XML to the Unitree scene_29dof.xml path.
e13 uses default model; e24 uses SM_IMU_MODEL=1 SM_WAIST_COUPLING=1 SM_WAIST_ROLL=range=0.72,damping=1.8.
e13 reference shanshenbu_combo4.csv is resampled 30Hz to 50Hz; goal conversion matches the bundled goal math within 4.77e-7.
All runs: friction 1.0, KP scale 1.0, extra mass 0; e24 comparison latency 0 or 20ms.
See policy-v2-results.json for measured values, hashes and exact conditions.
These are limited pretrained-policy playback checks, not retraining or the official 129,514-motion evaluation.

Example (paths are examples; obtain permitted artifacts from the upstream releases):
export G1_UNITREE_XML=/path/to/unitree_mujoco/unitree_robots/g1/scene_29dof.xml
python csv_to_goal.py --help
python capture_policy.py --repo /path/to/SuperMimic-Repro-Sim2Real --out /private/capture-e13 --ckpt e13i_0291000 --clip shanshenbu_combo4
MUJOCO_GL=egl python render_policy.py --repo /path/to/SuperMimic-Repro-Sim2Real --capture /private/capture-e13 --out /private/films --view reference
The renderer expects the production font path and ffmpeg h264_nvenc GPU1; adapt these to your machine. For CPU-only encoding replace the encoder section with libx264.
CSV conversion helper: csv_to_goal.py. Keep generated trajectory arrays private.