GR00T-N1.7-CubeStack-cotrain-abs

GR00T N1.7 (nvidia/GR00T-N1.7-3B) finetuned on a combined dataset of 16 synthetic Isaac Lab episodes plus 50 real-robot episodes (subset of unitreerobotics/G1_Dex3_BlockStacking_Dataset, remapped into the GR00T-LEAPP unitree_g1 hand-joint convention) of the Unitree G1 (Dex3 hands) cube-stacking task, head camera only (cam_left_high). Arm and hand actions are trained with absolute joint-position targets.

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

• Base model: nvidia/GR00T-N1.7-3B
• Training data: 66 episodes total (16 synth + 50 real, ~54.6k frames)
• Modality config: action groups left_arm, right_arm, left_hand, right_hand, all absolute
• Training steps: 6000, global batch size 8, lr 1e-4
• Final train_loss: ~0.219

Sibling checkpoints:

• GR00T-N1.7-CubeStack-synth-abs — synth-only, absolute
• GR00T-N1.7-CubeStack-synth-rel — synth-only, relative
• GR00T-N1.7-CubeStack-cotrain-rel — same data, relative arm actions

Status

Closed-loop evaluation in Isaac Lab sim (JointSpaceV2 env, head camera only): 0/5 successful stacking episodes. No measurable improvement over the synth-only checkpoint on this sample size. 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.