allros/GR00T-N1.7-grab-cube-so101

GR00T N1.7 — SO101 Grab Cube (dual camera)

Fine-tuned NVIDIA GR00T N1.7 on allros/lerobot_grab_cube1_trimmed (SO101 single arm, front + wrist cameras). Final checkpoint only (step 10000, merged weights, bfloat16, ~6.3GB).

Training

Item	Value
Base model	nvidia/GR00T-N1.7-3B
Embodiment	NEW_EMBODIMENT
Modality config	SO100 dual-cam (examples/SO100/so100_config.py)
Steps	10000
VLM	frozen (tune_llm=false, tune_visual=false)
DiT	frozen (tune_diffusion_model=false)
Action horizon	16

Inference (Isaac-GR00T)

python gr00t/eval/run_gr00t_server.py \
  --model-path <path-to-this-repo> \
  --embodiment-tag NEW_EMBODIMENT \
  --device cuda:0 --port 5555

LeIsaac client

python scripts/evaluation/policy_inference.py \
  --task=LeIsaac-SO101-GrabCube-Real-v0 \
  --policy_type=gr00tn1.6 \
  --policy_host=127.0.0.1 --policy_port=5555 \
  --policy_action_horizon=16 \
  --policy_language_instruction="Grab Cube" \
  --enable_cameras --device=cuda

Layout

• Repo root: final merged weights (model-*.safetensors, config.json).
• processor/, experiment_cfg/: required by Gr00tPolicy.
• meta/modality.json: LeRobot modality mapping (SO100 layout).