so101-chess-act-pick-wrist-v3

ACT pick policy for an SO-101 robot arm playing chess on a real board.

Wrist-only pick policy, third round. Kept because it was the baseline for v4.

  • Inputs: wrist camera, 6-D state. Output: joint targets, chunk 100. Run it with temporal ensembling (re-predict every frame, coefficient 0.01; --ensemble in so101-chess).
  • Trained on the first 205 episodes of Dev-devadath/so101-chess-pick-wrist-v4 (first round, d1/h1, starting position, full board ranks 1-2): 100k steps from scratch on 122 episodes, then two 30k-step fine-tunes as demos were added.

Results

  • Better than v1 and v2 on the starting squares, with neighbours around.
  • Weak in the middle of the board and on pieces the arm placed itself (10-20 mm off-centre): it had no demos there. Use v4.

Use

Made for the so101-chess setup: the policy starts at an IK hover 8 cm above a piece with the jaws open and only does descend, grasp, lift. It will not do anything useful from another start pose, and it expects the same jaw pads in the wrist view.

hf download Dev-devadath/so101-chess-act-pick-wrist-v3 --local-dir policies/so101-chess-act-pick-wrist-v3
.venv/bin/python chess/play.py --policy policies/so101-chess-act-pick-wrist-v3 --ensemble --skill 3

Built with lerobot at commit 5c43fa1c (0.5.2 dev).

Downloads last month
38
Safetensors
Model size
51.7M params
Tensor type
F32
·
Video Preview
loading

Dataset used to train Dev-devadath/so101-chess-act-pick-wrist-v3