Robotics
LeRobot
Safetensors
groot

GR00T N1.7 fully fine-tuned on toy pickup (toy_pickup)

nvidia/GR00T-N1.7-3B fine-tuned with LeRobot on the full toy pickup dataset (900 episodes).

Built on NVIDIA Cosmos.

Training

  • Full fine-tuning: all 3.14B parameters trained (vision encoder, language model and action head; tune_visual=true, tune_llm=true)
  • Embodiment tag new_embodiment; 2 cameras (front, wrist), 8-D state, 7-D action
  • Chunk 16, absolute actions, batch size 32, 30,000 steps (this checkpoint: step 30000)
  • AdamW lr 1e-4, cosine schedule with 5% warmup, fp32 weights with bf16 autocast, ColorJitter augmentation

Usage

lerobot-eval --policy.path=chihunn/groot_full_toy_pickup ...

License

Licensed by NVIDIA Corporation under the NVIDIA Open Model License. See LICENSE and NOTICE.

Downloads last month
-
Safetensors
Model size
3B params
Tensor type
F32
ยท
Video Preview
loading

Model tree for chihunn/groot_full_toy_pickup

Finetuned
(330)
this model