SuperMimicLite
G1 motion tracking with MimicLite and Uni-SuperMimic. Includes PPO and SuperMimic training, pretrained policies, ONNX export, and a fixed 228-motion evaluation.
Code · Data · Training · Evaluation
Demo
Coral: PPO. Green: SuperMimic. Gray: reference motion. Frames after the first failure are illustrative and excluded from the tracking score.
Results
| Policy | Method | Data | Body ↓ mm | Wrist ↓ mm | Root ↓ m | Completed / 228 |
|---|---|---|---|---|---|---|
| P-L | PPO | LaFAN40 | 40.22 | 62.61 | 1.479 | 120 |
| P-M | PPO | Mixed | 40.50 | 56.21 | 0.911 | 136 |
| S-L | SuperMimic | LaFAN40 | 43.22 | 65.55 | 0.737 | 182 |
| S-M | SuperMimic | Mixed | 32.47 | 42.37 | 0.511 | 198 |
Body and wrist errors average the Locomotion80 and Manipulation48 subset means equally. Root error uses Locomotion80. Completion counts all 228 motions before the first failure. See per-motion results and the evaluation protocol.
Policies
Final 32k policies and native checkpoints are hosted on Hugging Face.
| Policy | ONNX policy | Native checkpoint |
|---|---|---|
| P-L | P-L.onnx | P-L.pt |
| P-M | P-M.onnx | P-M.pt |
| S-L | S-L.onnx | S-L.pt |
| S-M | S-M.onnx | S-M.pt |
ONNX policies include normalization and, for SuperMimic, the encoder and decoder. PPO takes 535 policy and 304 command values; SuperMimic takes 990 history and 730 goal values. Both output 29 joint actions at 50 Hz.
Native checkpoints contain policy weights; S-L and S-M also contain the world model. They are final checkpoints without a complete optimizer resume state.
Download and evaluate
Requires Linux, Git, and uv. Model weights, 3,692 training motions, and the complete 228-motion evaluation set are public in this project's data hub. SONIC training motions are obtained separately under their original terms.
git clone https://github.com/han-xudong/SuperMimicLite.git
cd SuperMimicLite
uv sync --locked
uv run scripts/bootstrap.py
uv run scripts/fetch.py all --checkpoints
uv run scripts/smoke.py --out outputs/smoke-01
The downloader verifies pinned revisions and SHA256 hashes. ONNX files go to
models/, native checkpoints to checkpoints/, and evaluation motions to
data/motiondecode/. The robot asset is downloaded as well. To fetch weights
alone, use uv run scripts/fetch.py models --checkpoints.
For the full 228-motion benchmark:
uv run scripts/evaluate.py --out outputs/benchmark-01 --workers 4
uv run scripts/report.py --input outputs/benchmark-01/evaluation --out outputs/report-01
License
SuperMimicLite's original code is licensed under Apache-2.0. Third-party code and model or motion artifacts retain their own terms; see NOTICE and THIRD_PARTY.md.
Acknowledgments
Built on MimicLite, Uni-SuperMimic, and SuperMimic. See THIRD_PARTY.md for source revisions and licenses.
Upstream code, models, and source datasets retain their own licenses and terms. These results use one simulation seed; training and evaluation motions may overlap.
Training data includes Motion Data by Bones Studio. Use of the underlying dataset is subject to the BONES Motion Capture Dataset License Agreement.
