TangYishan/starvla-qwenoft-franka-robomimic-lift

StarVLA Franka QwenOFT (RoboMimic Lift, smoke)

Smoke checkpoint from reproducing examples/realRobots/Franka (no robot). Trained with StarVLA QwenOFT on RoboMimic "Lift" (real Franka, rtbenchmark `liftreal/ph`) converted to LeRobot v2.1 (180 train episodes, 10,351 frames).

1000 steps / batch 1 / Qwen3.5-0.8B backbone / MLP action head. This is an engineering smoke run, NOT a converged policy — do not use for real deployment.

Contract

• —action: 7D delta-EE [dx,dy,dz,droll,dpitch,dyaw,gripper], horizon 16
• —state: 6D eef_position(3) + eef_rotation(3, axis-angle) (include_state: false)
• —observation: 2 RGB views (baseview + egoview), 224×224
• —normalization: minmax (+ binary gripper); embodiment key `newembodiment`

Files

• —checkpoints/steps_1000_pytorch_model.pt — DeepSpeed checkpoint
• —config.yaml / config.full.yaml — training config
• —dataset_statistics.json — normalization statistics
• —summary.jsonl — step/loss summary

Caveats

• —Synthetic-observation validation only (no physical Franka). Actions carry no task semantics at this scale.