g1-atl-stomp-policy

G1 ATL Stomp — ONNX Whole-Body Control Policy

Fine-tuned SONIC whole-body control policy for the Unitree G1 humanoid (29 DOF) that performs ATL Stomp — a ~5-second street-dance routine with rapid stomp footwork, a single-leg hop phase, a deep squat with torso roll, and fast arm swings.

The stock SONIC base model fails this move: in the paired rollout below it loses balance and falls (5.06 s in), ending on the floor. The released policy completes the whole routine in free dynamics without falling and finishes standing at rest, on both the original-tempo clip and the 1.5× slowed variant it was trained on.

Results

Free-dynamics rollout on the same reference — the original-tempo clip plus a settle-to-stance tail (550 control steps @ 50 Hz) — every episode termination except time_out disabled, so nothing is masked by an episode reset:

The same policy on the 1.5× slowed reference plus a settle-to-stance tail runs all 544 control steps — root-height minimum 0.663 m, mean tracking error 96 mm — and ends the clip standing at rest.

See atl_stomp_before_after_en.mp4 in the companion GitHub repo (qjwdlwjdl/G1-ATL-Stomp) for a side-by-side render of both policies on the same reference.

Files

Exported with the official gear_sonic pipeline (eval_agent_trl.py +export_onnx_only=true).

Training

• Base model: nvidia/GEAR-SONIC (SONIC release checkpoint, universal-token whole-body controller).
• Method: PPO (TRL) motion tracking on a single street-dance clip, 2,048 parallel environments in Isaac Lab.
• Stage V8 (this release): 2,000 iterations, actor LR 4e-6, warm-started from stage V6 and trained on the repaired atl_stomp_v7 motion.

Three findings drove the final stage, each verified by rollout rather than by training reward:

1. Termination thresholds, not the policy, were the bottleneck. The stock training terminations (tracking/base_adaptive_strict_ori_foot_xyz) end an episode once the pelvis drifts 0.15 m or rotates 0.2 rad from the reference. On a single dynamic motion that fires the moment the policy starts to struggle, so PPO never sees — and never learns — the recovery region. Relaxing them (0.5 m / 1.2 rad / 0.4 m body, adaptive tightening off) moved the fall from 43 % to 90 % of the slowed clip within 500 iterations.
2. Long runs regress. At the stock 1e-5 learning rate the policy peaked around iteration 4,000 (90 % of the slowed clip) and had degraded back to 43 % by iteration 6,000. The final stage therefore warm-starts from the peak and refines at 4e-6.
3. The reference motion had a corrupted track. pose_aa[:, 0] (root rotation) disagreed with the clean root_rot quaternion by exactly 120° on 4 of 267 frames, while agreeing to within 0.05° everywhere else. Each of those frames teleports the reference (every tracked body jumps 10–56 cm in a single 20 ms step). atl_stomp_v7 rebuilds that column from root_rot.

Usage

The policy is compatible with the GR00T-WholeBodyControl deployment stack (Unitree G1, 29-DOF joint interface). Observations follow the SONIC universal-token policy interface (proprioception + motion-tracking command features at 50 Hz; actions are 29-DOF joint position targets).

Training data

Motion tracking on the in-house ATL Stomp motion clip — see the companion dataset repo (g1-atl-stomp-motion). Motion data captured in-house from a user-provided street-dance video; no third-party motion data used.

Credit

Motion Data by Bones Studio.

License

Fine-tuned weights released under Apache-2.0. Base model: NVIDIA GEAR-SONIC (Apache-2.0).