Downloads · 30 days
0
g1-taiji-form-sonic-onnx
oniichan521/g1-taiji-form-sonic-onnx
g1-taiji-form-sonic-onnx is a reinforcement learning model from oniichan521. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Fine-tuned SONIC whole-body control policy for the GHOST TRIAL 03 (SuperSONIC Challenge). Teaches a Unitree G1 humanoid a 25-second continuous Tai Chi / martial-arts routine (wide stances, low center-of-gravity drops,…
Access
Public
Updated Aug 14, 2026
Repo size
264 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.onnx264 MB · 100%
At a glance
- Task
- Reinforcement Learning
- License
- apache-2.0
- Access
- Public
- Created
- Aug 14, 2026
- Updated
- Aug 14, 2026
- SHA
- 96eddc26
Model card
From the Hugging Face model README
G1 Taiji Form — Fine-tuned SONIC Policy (ONNX)
Fine-tuned SONIC whole-body control policy for the GHOST TRIAL 03 (SuperSONIC Challenge). Teaches a Unitree G1 humanoid a 25-second continuous Tai Chi / martial-arts routine (wide stances, low center-of-gravity drops, lunges, weight shifts, torso rotations, large arm sweeps, and single-leg knee-lift balances).
Files
model_step_002000_g1.onnx— main policy (G1 robot)model_step_002000_encoder.onnx/model_step_002000_decoder.onnx— motion token encoder/decodermodel_step_002000_teleop.onnx— teleoperation variantmodel_step_002000_smpl.onnx— SMPL variant
Results (vs stock SONIC)
| Metric | Stock | Fine-tuned |
|---|---|---|
| mpjpe_g | 1242 mm | 189.1 mm |
| mpjpe_l | 55.8 mm | 34.3 mm |
| root drift | 1244.8 mm | 179.8 mm |
| Completion | fails in 0.24 s | full 25 s |
Training
Fine-tuned from nvidia/GEAR-SONIC release checkpoint: 4000 iters (single motion) + 3500 iters root-tracking refinement (anchor weight 1.0, std 0.2, lr 1e-5), dual L40S. Dataset: g1-taiji-form-dataset
Credit
Motion Data by Bones Studio
- Task
- Reinforcement Learning
- License
- apache-2.0
- Created
- Aug 14, 2026
- Updated
- Aug 14, 2026