NVIDIA previews GR00T N2, betting robots should predict video before they move

NVIDIA on March 16, 2026 previewed Isaac GR00T N2, the next generation of its robot foundation model, during chief executive Jensen Huang's keynote at the company's GTC conference. NVIDIA said the model uses a new world action model architecture taken from its DreamZero research and helps robots succeed at new tasks in new environments more than twice as often as leading vision language action (VLA) models. The company said GR00T N2 is slated to be available by the end of 2026.

The preview came with a wider set of releases. NVIDIA moved Isaac GR00T N1.7, a 3-billion-parameter open reasoning VLA model built for humanoids, into early access with commercial licensing, and said Humanoid, LG Electronics, NEURA Robotics and Noble Machines are adopting it. It also announced Cosmos 3, which it called the first world foundation model to combine synthetic world generation, vision reasoning and action simulation, and Isaac Lab 3.0 in early access, running on the new Newton 1.0 physics engine and the PhysX SDK. NVIDIA said GR00T N2 already ranks first on the MolmoSpaces and RoboArena leaderboards for generalist robot policies.

The research under N2 was posted on Feb. 17, 2026 as an arXiv preprint titled "World Action Models are Zero-shot Policies". Its authors include NVIDIA researchers Seonghyeon Ye and Joel Jang along with the company's robotics research leads Linxi "Jim" Fan and Yuke Zhu. DreamZero is a 14-billion-parameter autoregressive video diffusion model. An NVIDIA developer blog says it starts from the open Wan 2.1 image-to-video 14B model and that one transformer denoises video tokens and action tokens together, so the robot's next movements come out of the same pass that predicts what its cameras will see.

The preprint gives the numbers behind the claim. On AgiBot's G1 mobile two-armed robot, trained on about 500 hours of data from 22 environments, DreamZero reached 62.2% average task progress on tasks it had seen against 27.4% for pretrained VLA baselines, and 39.5% against 16.3% on unseen tasks. The baselines were NVIDIA's own GR00T N1.6 and Physical Intelligence's pi0.5. The authors cut inference latency from 5.7 seconds to 150 milliseconds, a 38-fold gain, on NVIDIA GB200 hardware through GPU parallelism, caching, CUDA graphs, NVFP4 quantisation and a faster "Flash" variant, which allowed closed-loop control at 7 Hz. They said weights, inference code and evaluation scripts are on GitHub.

Transfer between robots is the other headline result. According to the paper, adding 12 minutes of human egocentric video or 20 minutes of video from a YAM arm lifted performance on unseen tasks by more than 42% in relative terms, and the AgiBot-trained model adapted to the YAM robot with 30 minutes of unscripted play data while keeping its zero-shot ability.

GTC extended a programme NVIDIA laid out at CES on Jan. 5, 2026, when it released GR00T N1.6 as an open reasoning VLA with full-body control for humanoids, refreshed its Cosmos world models and began selling the Jetson T4000 module, a Blackwell part rated at 1,200 FP4 teraflops with 64 GB of memory, for $1,999 in 1,000-unit volumes. NVIDIA says 2 million developers use its robotics stack. The business logic is consistent across both events: the models are open or cheaply licensed, while training, simulation and onboard inference run on NVIDIA silicon.

The partner list at GTC showed how far that stack has spread. NVIDIA said FANUC, ABB Robotics, YASKAWA and KUKA, with a combined installed base of more than 2 million robots, are building its Omniverse libraries and Isaac simulation into virtual commissioning tools. AGIBOT is adopting GR00T N models for industrial humanoids, Skild AI and the Toyota Research Institute are among Cosmos 3 adopters, Hugging Face is folding Isaac and GR00T into its LeRobot framework, and Alibaba Cloud is integrating NVIDIA's full physical AI stack into its Platform for AI.

The bet is a technical one. VLA models such as pi0.5 and Google DeepMind's Gemini Robotics family adapt vision language models, trained mostly on images and text, to output motor commands. A world action model instead starts from a video generator that has absorbed how objects move and fall, and learns to predict actions alongside the next frames. Skild AI, which raised $1.4 billion at a valuation above $14 billion on Jan. 14, and Physical Intelligence are building their own general robot brains, and both compete with NVIDIA's models even as Skild uses NVIDIA tools.

Cost is the main open question. A later NVIDIA developer blog measured world action model inference at 590 and 800 milliseconds per action chunk against 190 milliseconds for pi0.5 in its own comparison, a 3 to 4 times slowdown, and reaching 7 Hz in the paper required data-centre GB200 hardware. The leaderboard rankings are NVIDIA's own statements, the preprint has not been peer reviewed, and GR00T N2 itself had not shipped by the end of the quarter.