GR00T-N1.7-CubeStack-synth-rel

GR00T N1.7 (nvidia/GR00T-N1.7-3B) finetuned on 16 synthetic Isaac Lab episodes of the Unitree G1 (Dex3 hands) cube-stacking task, head camera only (cam_left_high). Arm actions (left_arm, right_arm) are trained with relative targets (ActionRepresentation.RELATIVE, delta from the current observation, computed internally by GR00T's data pipeline). Hand actions stay absolute.

Part of the G1 Dex3 Cube Stacking Masterprojekt (Fichtl00/Masterprojekt_G1_CubeStack).

• Base model: nvidia/GR00T-N1.7-3B
• Training data: Fichtl00/Cube_Stacking_synth_groot_abs (16 episodes, GR00T-LEAPP unitree_g1 registry hand-joint convention; the underlying dataset stays absolute, only this checkpoint's modality config marks arm actions as relative)
• Training steps: 3000, global batch size 8, lr 1e-4
• Final train_loss: ~0.34
• Open-loop eval (vs. training trajectories): Avg MSE 0.0066, Avg MAE 0.054 (better than the absolute-action sibling)

Sibling checkpoints:

• GR00T-N1.7-CubeStack-synth-abs — same data, absolute arm actions
• GR00T-N1.7-CubeStack-cotrain-abs — + 50 real episodes, absolute
• GR00T-N1.7-CubeStack-cotrain-rel — + 50 real episodes, relative

Status

Closed-loop evaluation in Isaac Lab sim (JointSpaceV2 env, head camera only): 0/5 successful stacking episodes across multiple actuator stiffness/damping settings tested. Reach behavior is plausible (hand moves toward the correct cube) but grasp execution fails to produce a stable stacking outcome. Provided as-is for research/reproduction purposes.