bussing-whiteboard-act

bussing-whiteboard-act — BCE gripper head, stride-3, horizon 8

The whiteboard policy (frozen DINOv2-base encoder, 4-layer encoder / 7-layer decoder transformer, 142.3M params / 55.7M trainable) trained on the FR3 table-bussing recording.

Action space: 7 joint velocities (rad/s) + gripper position (8-d). Run at 5 Hz (min_period: 0.2). Gripper is the robot's RAW convention (1.0 = open); flip_gripper: False, asserted by the deploy adapter.

This revision (BCE-0.3)

The gripper dim is now a decisive BCE head instead of L1 regression: the demonstrator's carrying width reads raw ~0.47 — 0.03 under the 0.5 binarize line — and regression under the 80:20 open/closed imbalance parked carry predictions in the ambiguous 0.35–0.65 band (58% of held-out carry ticks). With the BCE head (pos_weight 4, loss weight 0.3) that band holds 2% of ticks; carry mean 0.08, open mean 0.99. config.json: gripper_head = "bce" — the shipped bussing_act/policy.py maps the logit back via 1 − sigmoid.

Held-out full-episode battery (latch close<0.5 / reopen >0.75×2, replan every tick, 0.2 s):

• grasps: 22/25 clean (3 close ≤0.6 s early, 0 arrive-closed, 0 never) — previous revision: 12/25 clean, 2 arrive-closed, 1 never
• carry-time false opens 2.5%, longest run 2.0 s (one confidently-wrong visual state in one episode; previous revision: 4-s hole in the same region)
• releases: 25/25 fire, median +0.4 s, none suppressed
• velocity cosine 0.628 (previous 0.608 — the mild BCE regularized the head)

Deploy with the paired adapter (whiteboard_policy.py, branch whiteboard-joint-velocity of sombit888/robot_evals): stateful gripper latch + temporal-order chunk execution + max_executed_chunk_size: 1 (E=1 beat E=2 head-to-head: 22/25 vs 19/25 clean grasps). Fallback: revision 230bf6b is the previous (regression-head) checkpoint.