groot-g1-sonic-grab-bottle-deploy-checkpoint-10000

GR00T N1.7 Unitree G1 SONIC Deploy Checkpoint

This is the deploy-only export of:

checkpoints/gr00t_n17_g1_sonic/grab_bottle_v2_4gpu_bs64_10k/checkpoint-10000

Resume-training artifacts such as optimizer state, RNG state, scheduler state, trainer state, and training-only configs are intentionally not included here.

Included Files

config.json
model.safetensors.index.json
model-00001-of-00002.safetensors
model-00002-of-00002.safetensors
processor_config.json
statistics.json
embodiment_id.json
experiment_cfg/final_model_config.json
experiment_cfg/final_processor_config.json
experiment_cfg/dataset_statistics.json

Training Summary

• Base model: nvidia/GR00T-N1.7-3B
• Embodiment: unitree_g1_sonic / UNITREE_G1_SONIC
• Dataset: cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v2
• Task: Unitree G1 right-hand bottle grasp
• Max steps: 10,000
• Precision: bf16
• Action keys: motion_token, left_hand_joints, right_hand_joints

Evaluation Summary

Open-loop checkpoint-10000 result:

• Avg MSE: 0.00029590
• Avg MAE: 0.01039191

MuJoCo / SONIC simulation result with lower body locked:

• Single bottle, grab bottle: 5 / 10 success
• Bottle + red apple, grab bottle: 3 / 10 success
• Bottle + red apple, grab red apple: 1 / 10 success
• Aggregate: 9 / 30 success

Real Unitree G1 deployment was not performed. The reported deployment success rates are simulation results only.

Usage

Load this directory as the GR00T policy checkpoint for SONIC VLA inference:

python gr00t/eval/service.py \
  --model-path /path/to/deploy_checkpoint_10000 \
  --embodiment-tag UNITREE_G1_SONIC \
  --server