G1 Wubuquan (Five-Step Fist) — SONIC fine-tuned policy (ONNX)
ONNX exports of the fine-tuned SONIC policy that makes a Unitree G1 humanoid perform a ~13.0-second Wubuquan-inspired martial-arts routine (Kimodo-generated): 5 stance types, straight punches and palm strikes, two direction changes, ~3.06 m net root displacement.
License: these weights are derivative models of NVIDIA's SONIC model — Licensed by NVIDIA Corporation under the NVIDIA Open Model License. See the LICENSE / NOTICE in the GitHub companion repository.
Performance
Apples-to-apples (no tracking termination; stock physically falls at ~4.5 s, so all models are compared on the same 0–4.5 s window):
| Metric | Stock SONIC (before) | Fine-tuned V1.1 (after) |
|---|---|---|
| mpjpe_g (global) | 198.1 mm | 76.0 mm (−62 %) |
| mpjpe_l (root-aligned) | 41.4 mm | 33.0 mm (−20 %) |
| mean root-position error | 190.1 mm | 64.1 mm (−66 %) |
Full-horizon (no tracking termination, all 649 samples): 133.1 mm
mpjpe_g / 46.8 mm mpjpe_l / 123.7 mm root error. The complete matrix
(matched-window, full-horizon, official terminated pipeline) is in the
GitHub repo's eval/metrics.md.
Files
model_step_001500_*— final policy (V1.1; 1500 more iterations, lr 1e-5, anchor std 0.2). Use this one. Five exports:decoder,encoder,g1,smpl,teleop.model_step_003000_*— alternate V1 policy trained on a shorter 5.14 s in-place backup clip (action_clip_martial_fixed, mpjpe_g 103.8 mm); reference only, not the submission model.
Details
- Training: SONIC/GEAR-SONIC, 2048 envs, L40S, 4000 iters (V1) then
+1500 iters root-focused refinement (V1.1 final). Exact run config in
the GitHub
run-config-v11release; full campaign history indocs/TUNING-NOTES.md(including two rejected experiments). - Companion dataset repo (motion CSVs + SONIC motion_lib pkl).
- GitHub: https://github.com/qjwdlwjdl/G1-Wubuquan-Five-Step-Fist
Credit
Motion Data by Bones Studio