so101-chess-act-pick-v1

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

The first pick policy. Works on a near-empty board, not in a real game.

  • Inputs: wrist + overhead cameras, 6-D state. Output: joint targets, chunk 100, played open loop (n_action_steps 100).
  • Trained from scratch on Dev-devadath/so101-chess-pick-v1, 100k steps, batch 8, lerobot ACT defaults (resnet18, lr 1e-5, amp), RTX 4000 SFF Ada. Loss 6.8 -> 0.025.

Results

  • 10 of 10 first-attempt picks across all piece types on a near-empty board, ranks 1-5, 3.3-4.1 s each.
  • Fails in a real game. With the black pieces on ranks 7-8, its grasps shift about half a square toward the arm, because it aims partly from the overhead view and never saw a full board. Superseded by the wrist-only policies.

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-v1 --local-dir policies/so101-chess-act-pick-v1
.venv/bin/python chess/play.py --policy policies/so101-chess-act-pick-v1 --skill 3

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

Downloads last month
35
Safetensors
Model size
51.7M params
Tensor type
F32
·
Video Preview
loading

Dataset used to train Dev-devadath/so101-chess-act-pick-v1