leepanic/gr00t-n1.6-pick-cube

GR00T N1.6 — Pick Cube (Unitree teleop)

Fine-tune of nvidia/GR00T-N1.6-3B on a Unitree pick_cube_demo teleop dataset.

Task instruction: pick up cube and place it on the green plate.

Embodiment

• State: 63 dims — left_arm(7) right_arm(7) left_ee(7) right_ee(7) body(35)
• Action: 31 dims — left_arm(7) right_arm(7) left_ee(7) right_ee(7) body(3)
• Camera: single ego_view at 640×480, 30 fps
• Action chunk: 16 steps, absolute non-EEF targets
• Embodiment tag: NEW_EMBODIMENT

Modality config: examples/pick_cube/pick_cube_config.py in the Isaac-GR00T fork used for training.

Training

base model	nvidia/GR00T-N1.6-3B
steps	10,000
global batch size	64
learning rate	1e-4 (warmup ratio 0.05)
weight decay	1e-5
GPUs	4
color jitter	brightness 0.3, contrast 0.4, saturation 0.5, hue 0.08

Usage

from gr00t.model.policy import Gr00tPolicy
from gr00t.data.embodiment_tags import EmbodimentTag

policy = Gr00tPolicy(
    model_path="leepanic/gr00t-n1.6-pick-cube",
    embodiment_tag=EmbodimentTag.NEW_EMBODIMENT,
    modality_config=pick_cube_config,   # see examples/pick_cube/pick_cube_config.py
)

Only the final model weights are published here — intermediate training checkpoints are not included.