NVIDIA Isaac GR00T N1

2025-03-18 · Publisher : NVIDIA · Domain : Robotics

At GTC 2025, NVIDIA releases Isaac GR00T N1, billed as the first open, fully customizable foundation model for humanoid robots: a dual-system VLA of about 2 billion parameters.

### What changes On March 18, 2025, at its GTC conference, NVIDIA released Isaac GR00T N1, which it billed as the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills. The model was available immediately, as the first of a family NVIDIA committed to pretrain and release. ### Specifications - **Architecture**: two systems inspired by human cognition. "System 2", a vision-language model, interprets the scene and instruction to plan; "System 1", a diffusion transformer, generates fluid motor actions in real time. Both are trained jointly end to end. - **Size**: about 2 billion parameters (GR00T-N1-2B on Hugging Face). - **Data**: real-robot trajectories, human videos and large volumes of synthetic data generated with Omniverse. - **Demo**: language-conditioned bimanual manipulation on the Fourier GR-1 humanoid. - **Related announcements**: the GR00T blueprint for synthetic data generation and Newton, an open physics engine developed with Google DeepMind and Disney Research. ### Why it matters By opening a base model for humanoids, NVIDIA gives robot makers a starting point without costly pretraining, while drawing them into its software stack and Jetson chips. ### Key takeaway GR00T N1 launched a family updated at a fast pace: N1.5 in May 2025, N1.6 in December 2025, N1.7 and the N2 preview in March 2026.

Tags : VLA, Humanoïdes, NVIDIA, Open Source, GTC