dodoleonarnci/uf850-groot-pickplace

UF850 GR00T-N1.7 fine-tune (pick & place)

Fine-tune of NVIDIA Isaac-GR00T N1.7-3B on UFactory 850 Meta-Quest teleoperation demos.

• Base model: nvidia/GR00T-N1.7-3B
• Embodiment: NEW_EMBODIMENT (7-DoF: joints j1..j6 + gripper)
• Task: "Pick up the cube. Place it on box. Pick up the cube again. Place it on table."
• Data: 101 LeRobot v2.1 episodes, base + wrist cameras (640x480 @ 30 fps)
• Training: 20,000 steps, global batch 32, action-head fine-tune (frozen VLM backbone)
• Final loss: 0.0237 (min 0.0198)

Inference checkpoint — optimizer state omitted. Load with the Isaac-GR00T stack using the NEW_EMBODIMENT modality config (single_arm 0:6 + gripper 6, base/wrist cameras).