Post-Training Isaac GR00T N1.5 for LeRobot SO-101 Arm
Hugging Face Blog
NVIDIA released Isaac GR00T N1.5, an updated foundation model for robot reasoning that can be fine-tuned for specific robot arms and tasks through post-training. The fine-tuning process requires approximately 25 gigabytes of VRAM and involves four steps: dataset preparation, model fine-tuning with 10,000 maximum training steps, open-loop evaluation, and deployment to physical hardware. Developers can now customize the model for diverse robot embodiments like the LeRobot SO-101 arm using the EmbodimentTag system instead of being limited to pre-trained configurations.
