ACT reach - dataset and trained checkpoint

Artifacts for the ACT visuomotor reach pipeline (see the README deep-dive).

• reach_act_dataset.tar.gz — the 40-episode bimanual-reach LeRobotDataset v3.0 (mp4 + parquet, ~6 MB). Load with LeRobotDataset(...) or retrain ACT directly from it.
• act_reach_checkpoint.tar.gz — a trained ACT policy (ResNet-18 + Transformer, use_vae=false), including the LeRobot pre/post normalization processors. Extract and load with ACTPolicy.from_pretrained(...). A representative single-seed run; retrain any seed with --seed N on the dataset above.
• act_reach_eval_3seeds.tar.gz — the raw evaluation data behind the README's headline numbers: per-episode reach errors for 3 independent training seeds (24 held-out rollouts each), the no-vision baseline, and a summary. Result: right 5.6 ± 0.6 cm, left 5.2 ± 0.3 cm, 69 % ± 9 % within 8 cm; no-vision baseline ~19.7 cm / 0 % — so the ~5 cm reach is the policy using the camera, not replaying a trajectory.

The skt_v3 MuJoCo model needed to render / roll out in the twin belongs to Rbotic/skate_teleop and is not redistributed here — fetch it via sim/make.py --clone.

Reproduction scripts: tools/skate_commander/examples/act_reach/. Debugging write-up: the ACT policy that reached for garbage.