xXFiEsTaDeAmOnXx/smolvla-openarmv2-cell

SmolVLA OpenArm v2 Cell

SmolVLA fine-tuned for the bimanual OpenArm v2 Cell pick-up task in MuJoCo simulation. Base model: lerobot/smolvla_base. Training data: enactic/openarm-2-cell-pick_up_cube_mujoco-lerobot.

Input cameras

Trained on three cameras (renamed to camera1–camera3):

policy input	dataset camera
observation.images.camera1	observation.images.ceiling
observation.images.camera2	observation.images.wrist_right
observation.images.camera3	observation.images.head_left

Task instruction: Pick up the cube. Action: 16-d joint positions (right arm, then left arm) at 30 Hz.

Evaluation

Evaluate with gym_openarmv2:

lerobot-eval \
  --policy.path=xXFiEsTaDeAmOnXx/smolvla-openarmv2-cell \
  --policy.device=cuda \
  --env.type=openarmv2 \
  --env.discover_packages_path=gym_openarmv2 \
  --env.cameras='["ceiling","wrist_right","head_left"]' \
  --rename_map='{"observation.images.ceiling":"observation.images.camera1","observation.images.wrist_right":"observation.images.camera2","observation.images.head_left":"observation.images.camera3"}' \
  --eval.n_episodes=10 \
  --eval.batch_size=1

Limitations

Fine-tuned on lift-only demonstrations (30 episodes). The placing phase of the cell task (set down inside the black tray) is not covered by the training data.