--- tags: [LeRobot, ACT, so101, robotics] datasets: [hrl-imerse/doug-stack-98-trimmed] --- # ACT block stacking, 80k steps (SO-101, wrist camera) ACT (chunk_size 100, batch 8), trained on `hrl-imerse/doug-stack-98-trimmed` for 170k steps (88 train / 10 val episodes). This is the 80k checkpoint. - Augmentation: colour jitter plus RandomAffine (±5°, ±5% shift). - Validation loss plateaued at about 0.57 after 10k steps. **Real-arm tests (2026-11-17, temporal ensembling coeff 0.01, one trial each):** | Checkpoint | Result | |---|---| | **80k** | **placed orange perfectly, then missed green**; the best of 20k/60k/80k/100k/170k | | 20k, 60k, 100k, 170k | 0 blocks | - Exec 15 was jerky and much worse than ensembling. - Grasps often missed by about 0.5 in on either side. Run with `lerobot-rollout ... --policy.n_action_steps=1 --policy.temporal_ensemble_coeff=0.01`.