act-ocbench-blue-cube-narrow

ACT policy (LeRobot implementation) for blue-cube pick-and-place in OCBench (MuJoCo, UR5e + Robotiq 2F-85) from three 224x224 RGB cameras and 7-D proprioception, trained on the 217 episodes of Autobrik/ocbench-blue-cube-narrow-delta. Actions are joint and gripper-driver deltas (target minus current state). Code, protocol and full results: https://github.com/BrikHMP18/act-weight-editing (see TRIAL1_RESULTS.md).

The repository root is the step-80,000 checkpoint, selected by blue-cube success on diagnostic seeds (20/20). All saved steps are under checkpoints/<step>/. Selection by validation loss would have picked step 20,000, which fails (0/10).

Task setting (narrow pilot): cube yaw within +-10 degrees, cube in x 0.30-0.55 m and y 0.15-0.35 m, fixed home arm pose, fixed goal (0.425, 0). Execute 50 actions per chunk (no temporal ensembling); this matters a lot (0% to 90% across chunk sizes).

Final test (seeds 20000-20049, strict success)	Success	95% CI
Blue 6 cm cube	44/50 = 88%	76-94%
Red 6 cm cube	45/50 = 90%	79-96%
Red 9 cm block	42/50 = 84%	71-92%

Edited checkpoints (edits/)

Folder	Edit	Red final test
edits/conv1-color-fold	red->blue RGB map folded into conv1 + bn1 (9,472 weights)	44/50 (blue 44/50)
edits/bn1-channel-color-k8	per-channel bn1 color correction (16 numbers)	diagnostic only: 18/20 vs 19/20
edits/control-random-conv1	random edit, same tensors and size (control)	0/50

Each folder has the full checkpoint and an edit.json with the per-tensor diff.

Usage

hf download Autobrik/act-ocbench-blue-cube-narrow --exclude "checkpoints/*" "edits/*" --local-dir ckpt
uv run act-rollout --checkpoint ckpt --action-space delta --n-action-steps 50 \
  --cube-yaw-range 10 --cube-xy-bounds 0.30 0.15 0.55 0.35 --fixed-arm \
  --device cuda --color red --seed 20000 --episodes 50