NVIDIA Isaac GR00T N1

Isaac GR00T N1, announced by NVIDIA at GTC on March 18, 2025, was pitched as the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills. It's a vision-language-action model of roughly 2 billion parameters with a dual-system architecture inspired by human cognition: a slow-thinking System 2 (a vision-language model that reasons about the environment and plans actions) paired with a fast-thinking System 1 (a flow-matching diffusion transformer that turns plans into precise, continuous motor commands). System 1 was trained on human demonstration data plus massive synthetic data from NVIDIA Omniverse.

NVIDIA released the weights, training data, and benchmark tasks on Hugging Face and GitHub, and has since iterated fast: N1.5 (June 2025), N1.6 (September 2025), and N1.7 (April 2026), which moved to a commercial-friendly NVIDIA Open Model License. The original N1 shipped under non-commercial terms, so check the license of the exact checkpoint you use. GR00T N1 generalizes across grasping, single- and dual-arm manipulation, and multi-step material-handling tasks — provider-claimed capabilities, since NVIDIA published no standardized benchmark scores for it.

• No standardized public benchmark scores published for GR00T N1
• NVIDIA-reported: generalization across grasping, moving objects with one or both arms, and multi-step manipulation tasks (provider claim)