mega-joanne/gr00t-n17-burger-bi-so-20k

GR00T N1.7 Burger bi_so_follower Fine-tuned Checkpoint

This checkpoint was fine-tuned from NVIDIA GR00T N1.7 on the sdhws/burger_merged_01 LeRobot dataset.

Summary

• Base model: nvidia/GR00T-N1.7-3B
• Dataset: sdhws/burger_merged_01
• Robot type: bi_so_follower
• Hardware: NVIDIA DGX Spark / NVIDIA GB10
• Fine-tuning steps: 20,000
• Global batch size: 8
• Final checkpoint: checkpoint-20000

Dataset / Embodiment Mapping

State/action dimension: 12

• left_arm: indices 0-4
• left_gripper: index 5
• right_arm: indices 6-10
• right_gripper: index 11

Cameras:

• observation.images.left_top
• observation.images.left_left
• observation.images.right_right

Offline Evaluation

Trajectory 0-9 comparison:

Checkpoint	Average MSE	Average MAE
checkpoint-100	1122.724854	25.430536
checkpoint-20000	31.217651	3.145292

Usage Note

This checkpoint requires registering the custom NEW_EMBODIMENT modality config before inference.

See:

custom_config/burger_bi_so_config.py
custom_config/modality.json

Safety Note

Real robot deployment still requires:

• action range checks
• joint limit clamps
• velocity/rate limits
• low-speed dry-run testing
• emergency stop / manual override