g1-inspire-nut-in-box-n16-rel
GR00T N1.6 fine-tune for the Unitree G1 + Inspire "nut in box" tweezer task
("pick up the tweezers and put the nut in the box").
- Dataset:
MLeggiero/g1-inspire-nut-in-box-twist2, converted to LeRobot
v2.1 and success-gated to 29 episodes (~58k frames @ 60 fps).
- Cameras: head (1280x720) + both wrist fisheyes (1920x1080).
- State (17-D): waist 3 + left arm 7 + right arm 7.
- Actions (27-D contract): arms 14 + right-hand Inspire registers 6 +
root height / lin-vel-xy / yaw-rate 4 + waist 3.
- Action mode: RELATIVE (
use_relative_action=true, chunk-anchored;
serving must decode back to absolute against the input state — GR00T's
own processor does this in get_action). Horizon 30.
- Training: 10,000 steps, effective batch 256, bf16, single B200;
final train_loss 0.040.
- Open-loop gate (3 episodes, 18 probe frames): arms corr 1.00 /
RMS 0.026 rad, waist corr 1.00, hand corr 1.00 / RMS ~12 registers
(0-1000 scale), closure ramp tracked. Checkpoint step 10000.
Sibling runs (same data/contract): an absolute-action arm and a
tune-visual arm are training; pi0.5 arms (ours + official recipe)
train separately.
