g1-inspire-nut-in-box-pi05
Ο0.5 (3.6B) full fine-tune for the Unitree G1 + Inspire "nut in box"
tweezer task ("pick up the tweezers and put the nut in the box") β our
recipe, trained with openpi's PyTorch path (config pi05_g1_nut_full).
- Dataset:
MLeggiero/g1-inspire-nut-in-box-twist2, LeRobot v2.1 conversion, success-gated 29 episodes (~58k frames @ 60 fps). - Cameras: head (1280x720) + both wrist fisheyes (1920x1080), full-frame (openpi resizes to 224 internally).
- State: 34-D body+imu, delivered as Ο0.5's discrete prompt tokens, with our fork's state-token dropout p=0.8 at train time (inference always sees state) β the anti-shortcut regulariser.
- Actions (27-D contract): arms 14 + right-hand Inspire registers 6 +
root height / lin-vel-xy / yaw-rate 4 + waist 3 β ABSOLUTE
(
use_delta_actions=False), horizon 30, quantile (q01/q99) norm. Norm stats ship insideassets/(the checkpoint's own). - Training: 20,000 steps, batch 32, AdamW (Ξ²2 0.95), cosine warmup 1k β peak 2.5e-5 (30k decay horizon β final LR ~8e-6), bf16, single B200.
- Open-loop gate (3 episodes, 15 probe frames, vs dataset GT): arms RMS 0.004/0.010 rad, hand ~2.0 registers (0-1000 scale), waist 0.002 rad, root 0.011 β corr 1.00 on every group; 4-5x tighter than the GR00T N1.6 twins trained on the same capture. Monotonic across 16k/18k/20k.
- This is the final step-20000 checkpoint (weights + norm-stats assets + metadata; no optimizer state).
Serving: load with openpi's create_trained_policy(get_config( "pi05_g1_nut_full"), <this dir>) β pass the bundled assets' norm stats
if your config's repo_id differs from the training one.
Siblings: g1-inspire-nut-in-box-n16-abs / -n16-rel (GR00T N1.6 twins
on the same contract).