starvla-qwenoft-g1-sonic

StarVLA QwenOFT — Unitree G1 WholeBody (SONIC 78D action)

Smoke checkpoint for the StarVLA examples/realRobots/UnitreeG1_WholeBody workflow.

What this is

• Model: QwenOFT — Qwen3.5-0.8B backbone + L1-regression MLP action head
• Action contract: SONIC-style 78D = 64D motion_token + 7D left_hand_joints + 7D right_hand_joints
• State: 72D proprioception = 43D observation.state + 14D observation.eef_state + 4D root_orientation + 3D projected_gravity + 4D cpp_rotation_offset + 4D init_base_quat
• Normalization: q99 (dataset_statistics.json; unnorm_key = new_embodiment)
• Image: single ego view, 224×224
• Horizon: 8

Scope / caveats

• 1000-step SMOKE run — not converged (final L1 loss ≈ 0.60). Do not expect task-level performance.
• Trained on the public cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v11 dataset (355 episodes), not the private test_sonic example data.
• Validates the data → train → policy-server → 78D-action-split engineering chain. It does not imply real-robot closed-loop control or task success (no physical G1 / SONIC / WBC used).

Files

• checkpoints/steps_1000_pytorch_model.pt — StarVLA checkpoint (2.24 GB)
• config.yaml / config.full.yaml — model + training config
• dataset_statistics.json — q99 normalization stats (required for serving)
• summary.jsonl — training checkpoint-save log

Serving (StarVLA)

Requires the StarVLA repo and the Qwen3.5-0.8B backbone (base_vlm in config.yaml):

cd starVLA
export PYTHONPATH=$PWD:${PYTHONPATH}
CUDA_VISIBLE_DEVICES=0 python deployment/model_server/server_policy.py \
  --ckpt_path <this_repo>/checkpoints/steps_1000_pytorch_model.pt \
  --port 5694 \
  --use_bf16

Reproduction

Training config / scripts live in the StarVLA example: examples/realRobots/UnitreeG1_WholeBody/step2_training/train_files/.

• Dataset: cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v11 (LeRobot v2.1)
• Backbone: Qwen/Qwen3.5-0.8B
• dataset_statistics.json top-level key new_embodiment = EmbodimentTag.NEW_EMBODIMENT