PID0930/gr00t-n17-g1-inspire-broccoli-tray-20k

GR00T N1.7 G1 Inspire Broccoli Tray 20k

Fine-tuned GR00T N1.7 checkpoint for a Unitree G1 robot with Inspire DFX56 five-finger hands.

Task instruction:

Pick up the broccoli plush with the right hand and place it on the red tray.

Training Summary

• Base model: nvidia/GR00T-N1.7-3B
• Dataset: PID0930/g1-inspire-pick-cube-gr00t-lerobot
• Embodiment tag: new_embodiment
• Training steps: 20,000
• Final train loss: 0.060223662567138674
• W&B run: gr00t-n17-g1-inspire/g1-inspire-broccoli-tray-20k
• W&B URL: https://wandb.ai/yuto083/gr00t-n17-g1-inspire/runs/if2wc99c

Files

This repository contains the final model artifacts needed for inference:

• config.json
• model-00001-of-00002.safetensors
• model-00002-of-00002.safetensors
• model.safetensors.index.json
• processor_config.json
• statistics.json
• embodiment_id.json
• training_args.bin
• experiment_cfg/

Optimizer states, RNG states, scheduler states, and intermediate checkpoint-* directories are intentionally excluded.

Notes

The dataset was converted from Unitree G1/Inspire JSON and JPEG recordings into LeRobot v2 format. It contains 9 usable episodes, 3,979 frames at 30 FPS, and one ego-view camera stream. State and action vectors have 26 dimensions: left arm 7, right arm 7, left hand 6, and right hand 6.