2chan/coffee-bimanual-smolvla

Bimanual SO-101 coffee SmolVLA

SmolVLA was fine-tuned for a two-arm SO-101 coffee preparation demonstration.

Artifact

• Branch: main
• Base model: lerobot/smolvla_base at revision d9f33c94a60fb382c90dea2164c96845bd955e28
• Offline best: step 10000, eval loss 0.1484
• Dataset revision: 6cb48fc7215028d0bf584400436ce2c558f52e76
• Selection criterion: minimum held-out episode evaluation loss; ties prefer the earlier step.

Use main for the final training step and best-eval-loss for the automatically selected model.

lerobot-rollout \
  --policy.path=<THIS_REPO> \
  --policy.pretrained_revision=main \
  --robot.type=bi_so_follower \
  --task="Hold the cup with the right arm. With the left arm, add powder to the cup, pour water into it, and stir with a stirrer."

Inputs

• observation.images.left_wrist

• observation.images.right_wrist

• observation.images.top

• observation.state: 12 absolute SO-101 joint/gripper positions

• Language instruction: Hold the cup with the right arm. With the left arm, add powder to the cup, pour water into it, and stir with a stirrer.

Output order

1. left_shoulder_pan.pos
2. left_shoulder_lift.pos
3. left_elbow_flex.pos
4. left_wrist_flex.pos
5. left_wrist_roll.pos
6. left_gripper.pos
7. right_shoulder_pan.pos
8. right_shoulder_lift.pos
9. right_elbow_flex.pos
10. right_wrist_flex.pos
11. right_wrist_roll.pos
12. right_gripper.pos

The machine-readable camera, named-joint, task, and revision mapping is stored in training_contract.json and is checked before real-robot rollout.

Access and license note

The repository uses manual gated access. The base model is Apache-2.0, while the linked dataset currently has no declared license; this derived model is therefore marked other until the dataset owner documents redistribution terms.