NVIDIA · 2025
GR00T N1: an open foundation model for humanoids
GR00T N1 is NVIDIA's open vision-language-action model for humanoid robots. It ships the dual-system design — slow reasoning, fast reflexes — as a foundation others can adapt to their own hardware, backed by NVIDIA's simulation and synthetic-data stack.
What it is
Dual-system, out of the box
GR00T N1 pairs a vision-language model (System 2) that reasons about the scene and goal with a diffusion transformer (System 1) that produces high-rate motor commands. It is the standard VLA architecture, packaged as an open model for humanoid robots.
Because it is open, robot developers can fine-tune GR00T N1 onto their own humanoid rather than training a controller from zero — the same "start from a foundation" move that reshaped language AI.
How it is trained
Real data is scarce — so NVIDIA makes more
The hardest constraint for any VLA is data: real robot demonstrations are expensive to collect. GR00T N1 addresses this with a blend of real robot data, human video, and large-scale synthetic demonstrations generated in NVIDIA's Isaac and Cosmos stack. Simulation and synthetic data are how the model reaches scale without a warehouse of teleoperators.
GR00T N1 gives a humanoid its skills. It does not give it knowledge of a specific home or worksite — that context is the knowledge layer benned provides on top.
FAQ
Common questions
What is GR00T N1?
An open foundation vision-language-action model for humanoid robots from NVIDIA (2025, Project GR00T). It uses a dual-system design — a vision-language model for reasoning plus a diffusion transformer for fast motor control — trained on real, synthetic, and human-video data.
Is GR00T N1 open?
Yes. NVIDIA released it as an open foundation model so developers can adapt it to their own humanoid hardware, paired with the Isaac simulation and Cosmos synthetic-data tools.
Last updated: July 2026