NVIDIA Isaac GR00T N1.5 Post-Training on LeRobot SO-101 Using EmbodimentTag

AI Impact Summary

NVIDIA releases GR00T N1.5, the first major update to Isaac GR00T, a cross-embodiment, multimodal foundation model for humanoid robot reasoning and manipulation. The update enables post-training for specific embodiments (e.g., LeRobot SO-101) via the EmbodimentTag system, with a full workflow from dataset preparation to deployment. Teams should plan for significant local compute (default fine-tuning around 25 GB VRAM) and use open-source data sources such as the so101-table-cleanup dataset on Hugging Face as part of their fine-tuning process. This lowers the barrier to tailoring advanced robot policies to custom hardware, enabling faster time-to-value for embodied AI projects, but requires careful resource planning and validation on target hardware.