NVIDIA Isaac GR00T N2

NVIDIA's next-generation robot foundation model, previewed for 2026, moving from a VLA design to a world-action-model architecture.

GR00T N2 is the next-generation general-purpose robot foundation model NVIDIA previewed at GTC 2026 on March 16, 2026, succeeding the GR00T N1 series (N1, N1.5, N1.6, N1.7). The N1 series followed a VLA design: a vision-language model understands the scene and instruction, and a diffusion-style action head outputs actions. According to NVIDIA's announcement, N2 is “based on DreamZero research” and switches to a world-action model (WAM) architecture: built on a pretrained video generation model, it predicts future frames and robot actions at the same time, learning physical dynamics from video rather than just semantics. NVIDIA says it succeeds at new tasks in new environments more than twice as often as leading VLAs, and at launch it ranked first on the MolmoSpaces and RoboArena general-policy benchmarks. NVIDIA planned to release it before the end of 2026; as of September 2026, the newest GR00T weights NVIDIA had published on Hugging Face were still from the N1.7 series.

ExampleIts research foundation, DreamZero, is built on a 14-billion-parameter autoregressive video diffusion model that, once optimized, can control a robot in real time closed-loop at 7Hz; adapting to a new robot takes only about 30 minutes of play data.

Also called

Isaac GR00T N2, GR00T N2

Related

NVIDIA Isaac GR00T N1 · DreamZero · World Action Model · Vision-Language-Action Model · RoboArena · NVIDIA

Sources

NVIDIA and Global Robotics Leaders Take Physical AI to the Real World (NVIDIA Newsroom, 2026-03-16)
World Action Models are Zero-shot Policies (DreamZero, arXiv 2602.15922)
NVIDIA/Isaac-GR00T GitHub 仓库 (Chinese)

As of

2026-09