SO-101 bimanual policies (π0.7 recipe on the π0.5 backbone), full precision
Checkpoints as saved by training (model.pt state dict plus config.json, stats.json, extra.json), for
the pi_model code in the project repo (docs/INFERENCE.md). The backbone is LeRobot's PI05Pytorch
(lerobot/pi05_base); the PaliGemma tokenizer is not included (get it from google/paligemma-3b-pt-224).
| folder | strict success on 30 fixed seeds per task (pick / bin / handover / two_bins) |
|---|---|
pi05_dagger2_step7000 |
28 / 28 / 12 / 18 = 86/120 (best two-arm, best two_bins) |
pi05_v1_step20000 |
26 / 26 / 19 / 12 = 83/120 (best handover) |
pi05_single_step6000 |
single-arm pick_place 28/30 |
Evaluated closed loop in the matching MuJoCo scene with oracle subtask prompts, 25 actions per chunk, no delay.