NVIDIA Isaac GR00T N2 (preview)

What it is

Isaac GR00T N2 is a robot foundation model that NVIDIA previewed as a next-generation entry in its GR00T line. Jensen Huang, NVIDIA's founder and CEO, showed the preview during his GTC keynote in March 2026 ^[1]. NVIDIA describes the model as building on DreamZero research and as using a new world action model architecture ^[1]. A robot foundation model is a general-purpose model meant to be adapted across tasks and embodiments rather than hand-built for one narrow job. GR00T N2 is presented in that spirit: a base model for generalist robot behavior.

How it is described

NVIDIA's account centers on the architecture. The company calls GR00T N2 a next-generation robot foundation model based on DreamZero research, and says it is built on a new world action model architecture ^[1]. In general, a world model tries to represent how a scene or situation may change; an action model turns such representations, together with the task, into robot actions. NVIDIA's naming places GR00T N2 in that combined direction. The company contrasts its results with vision-language-action models, usually shortened to VLA: models that take visual and language input and produce actions. NVIDIA says GR00T N2 helps robots succeed at new tasks in new environments more than twice as often as leading VLA models ^[1]. That comparison is about generalization, meaning the ability to handle tasks and settings beyond a narrow, fixed setup, rather than about repeating one demonstration.

Results and benchmarks

At the preview, NVIDIA said GR00T N2 ranks first on MolmoSpaces and RoboArena for generalist robot policies ^[1]. These names place the claim in the evaluation landscape for general-purpose robot control: a policy is the rule or model that maps what a robot sees to what it does. The ranking is one of the concrete claims NVIDIA makes for the previewed model. The broader performance claim is the comparison with leading VLA models on new tasks in new environments, where NVIDIA says the model succeeds more than twice as often ^[1].

Position in the line

The preview was announced alongside GR00T N1.7 and the Cosmos 3 world foundation model on 16 March 2026 ^[1]. That grouping shows how NVIDIA is presenting GR00T N2: as part of a wider set of robot and world models rather than as an isolated item. GR00T N1.7 is the adjacent GR00T model named in the announcement, and Cosmos 3 is described as a world foundation model ^[1]. For readers following robot learning, the interesting part is the combination of a generalist policy model with world-model research, which is the direction NVIDIA says it is taking with N2.

Why it matters

Generalist robot models matter because new-task and new-environment success is central to generalist robotics. A model that can carry over to unfamiliar tasks reduces the need to engineer every behavior separately. NVIDIA's claim of more than double the success rate of leading VLA models is therefore the headline result, while the first-place rankings on MolmoSpaces and RoboArena are supporting evidence in the company's account ^[1]. GR00T N2 is slated to be available by the end of the year ^[1]. The preview is thus a statement of direction: a world action model architecture, benchmark claims, and a planned release.