GR00T-N1.7-g1-apple2plate
GR00T N1.7 policy fine-tuned on a Unitree G1 humanoid for apple-to-plate pick-and-place (red or green apple), including correction/redirect demonstrations.
Robot
- Embodiment:
Unitree_G1_Dex1_UpperBody_2Cam(Dex1 parallel grippers) - State / action: 16-D (left arm 0:7, right arm 7:14, left gripper 14:15, right gripper 15:16)
- Cameras:
cam_head,cam_right_wrist@ 30 Hz
Training
- Base model:
nvidia/GR00T-N1.7-3B - Dataset:
tysyuvraj/g1-apple2plate-lerobot-v21(186 episodes, LeRobot v2.1 — training format) |tysyuvraj/g1-apple2plate(same data, LeRobot v3.0 — deployment/eval format) - Steps: 30000 (final) | batch 32 | lr 1e-4 | bf16 | single GPU
- Embodiment tag:
new_embodiment, modality:examples/g1_gripper/g1_gripper_config.py
Task instructions (use verbatim)
pick up the red apple and place it on the plate.pick up the green apple and place it on the plate.
Serve with the GR00T inference server and send one of the exact strings above as the language instruction.