RooibosT/gr00t-n1.7-g1-dex1-bct-relarm-aug-30hz-h40
GR00T N1.7 — Unitree G1 + Dex1, Building Children's Table (30 Hz, relative arms, augmented state)
Action-expert fine-tune of nvidia/GR00T-N1.7-3B for a Unitree G1 with Dex1 grippers assembling a children's table. Selected by open-loop scan over 6 checkpoints, not by eval_loss (see Checkpoint selection below).
Inputs
Outputs
Chunk of 40 steps (1.33 s at 30 Hz) over 19 dims:
Legs are observed but not commanded — locomotion stays with the whole-body controller.
State conventions (must be reproduced exactly at deployment)
base_gravity— gravity direction in the pelvis frame, from the IMU quaternion(w,x,y,z):-[2(xz - wy), 2(yz + wx), 1 - 2(x² + y²)]. Yaw-invariant, unlike the raw quaternion whose w/z components are ~all session heading (yaw std 109° vs roll 0.6°).left_eef/right_eef— wrist pose recomputed by forward kinematics from the arm joints in the same row, not the recorded end-effector state (which lagged the joints by 10 frames / 333 ms). Convention:g1_body29_hand14.urdf, waist joints zeroed (torso frame), wrist_yaw origin translated +0.05 m along local x, orientation as extrinsic xyz Euler.- Root x/y/z are deliberately absent: x,y vary ~90x more between clips than within one (a clip-ID shortcut) and z is a deterministic function of the leg joints already in state.
A mismatch in these conventions silently corrupts 15 of 46 input dims. state_dropout=0.2 makes the model robust to missing state, not to wrong state.
Open-loop accuracy (66 held-out episodes, 529 windows, unnormalized units)
Error grows with horizon, so the full-chunk figure is not the deployment figure — execute the head of the chunk and re-infer. A hold-current-pose baseline gives 1.13° at h1 and 5.72° over the full chunk. For scale, the arms move 10.2° over a 40-step chunk.
Training
Checkpoint selection
eval_loss is the flow-matching regression objective at random noise levels; across five runs it rose after ~7.5k steps while open-loop action accuracy kept improving. It is not a model-selection signal here. Checkpoints were instead scanned with a fixed-seed open-loop evaluation on the held-out split.
A longer schedule was tried and rejected: 30,000 steps at effective batch 256 saturated by ~20k and finished slightly worse than this model (full-chunk MSE 0.0299 vs 0.0291, gripper 0.204 vs 0.198), so roughly 5-6 epochs is enough for this dataset.