NVIDIA Isaac GR00T N1.5 enables post-training for LeRobot SO-101 arm via EmbodimentTag customization

AI Impact Summary

Isaac GR00T N1.5 delivers a major upgrade to the world’s first open foundation model for generalized humanoid robot reasoning, with a post-training workflow that tunes the model to a specific embodiment (SO-101) using teleoperation data. The EmbodimentTag system enables cross-embodiment customization, enabling developers to tailor perception and manipulation to the LeRobot platform, potentially reducing time-to-value for robotics pilots. The release provides end-to-end steps (dataset prep, fine-tuning, evaluation, deployment) and uses open datasets from HuggingFace, but teams must provision GPUs with around 25 GB VRAM for default training and manage data pipelines accordingly.