act_carton_middle_labels_2k

The selected checkpoint of the middle-labels fine-tune. ACT, 2,000 steps from act_carton_phase_c_50k, on middle-pose episodes with hand-authored recovery annotations.

Task: dual-view, PressureVision-supervised paper-carton pick with an SO-101.

Selected on an offline comparison over held-out middle episodes 8, 9 and 12 (654 frames):

Policy	Body MAE / p95	Open-phase false close
parent aug-50k	1.750 / 3.641	0/172
fine-tune 1k	1.252 / 2.771	12/172
fine-tune 2k	1.355 / 2.951	2/172
fine-tune 3k	1.355 / 2.817	3/172
fine-tune 4k	1.404 / 2.866	0/172
fine-tune 5k	1.428 / 2.986	3/172

It was chosen as a bounded physical candidate, not a universal offline winner — 1k has the lowest MAE but closes falsely six times as often.

This is not approval for autonomous deployment. On the arm it produced real lifts under the fixed-middle contract, and also no-lift, slip and no-release outcomes. A from-scratch control on the same data is a documented NO-GO: it falsely closed on all 172 held-out open-phase frames through 35k steps.

There is no train_config.json here; the training script writes only the policy and its processors. The recipe, the episode selection and the recovery labels are in training/train_middle_labeled_act.py, and the evaluation above is in training/TRAINING_HANDOFF.md.

Data: hand_tracking_pv_carton_middle_standard episodes 0-7, 10, 11 + hand_tracking_pv_carton_phase_b episodes 1, 9, 14, 20, 26.