NVIDIA has introduced GR00T 2, an open world action model for robots, at the Conference on Robot Learning in Austin, Texas. The announcement, published by NVIDIA on 9 November 2026, presents GR00T 2 as a model that learns dexterous manipulation by predicting what a robot will see next and what it should do next, together. It arrives one day before Google DeepMind’s Gemini Robotics 2.5 launch, and the two releases frame the same race from opposite ends: DeepMind proving a model inside a partner’s factory, NVIDIA opening the model and the training stack around it.

What NVIDIA confirmed in its own announcement

GR00T 2 is built on NVIDIA Cosmos and trained on diverse robot and human data, including 100,000 hours of NVIDIA’s open source human egocentric data: video recorded from a person’s point of view, which lets a model learn the structure of human manipulation before it ever touches a robot body. NVIDIA describes the category as a world action model, distinct from a vision-language-action model because the system predicts future video frames alongside the actions, then executes the actions it has just imagined.

NVIDIA’s reported results, from its own evaluations, are these. Out of the box, GR00T 2 completed 40 of 50 evaluated tasks at least once without any task-specific demonstrations. With one demonstration per task, it reached 71 percent average success across the 50 tasks. With five demonstrations per task, it reached 95 percent. NVIDIA also says the model currently ranks among the top three models across the vision-language-action and world action categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. Those are vendor-reported figures from NVIDIA’s own testing, with the selection method for the 50 tasks described in NVIDIA’s research material; they are not independently verified results.

The announcement also covers the surrounding stack rather than the model alone: Isaac Lab for robot learning in simulation, new benchmarks, and agentic workflows in which AI agents use GR00T 2’s API and supporting tools to run skills in sequence and recover from errors. NVIDIA names early users: Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, Siemens is evaluating it across industrial use cases, and MicroAGI is using a reinforcement-learning post-trained version for Ethernet cable swapping with a robotic arm.

Open model, with the release details still to come

NVIDIA calls GR00T 2 open, and the direction of travel matters for the robotics ecosystem: an open foundation model with skill transfer between different robot bodies lowers the cost of starting a serious manipulation programme. The caution for teams planning around it is timing. Coverage of NVIDIA’s research pages notes that code is listed as arriving in December, and no license terms for the weights had been published alongside the announcement. Treat GR00T 2 today as an announced platform to evaluate and prepare data for, not as a downloadable asset your roadmap can already depend on.

Why two robotics launches in two days matter

DeepMind’s Gemini Robotics 2.5, covered in our report on the Gemini Robotics 2.5 Mercedes factory pilot, is a closed model rolling out to selected partners with a factory proof of concept attached. GR00T 2 is an open model with benchmark results and industrial evaluators attached. Manufacturers and robotics startups now have both patterns in front of them at once: rent capability from a frontier lab, or build on an open stack and own the adaptation work. That choice, and the licensing fine print still to come from NVIDIA, will shape more purchasing decisions in 2027 than any single benchmark score.

The industrial backdrop is the data centre build-out itself, the same wave behind deals such as the photonics acquisition in our Marvell and Lumentum analysis, and the scrutiny on NVIDIA’s leadership covered in our Nvidia board investigation report. Physical AI is where that compute story meets the factory floor, and this week’s two launches show both of its leading suppliers moving at once.

What to watch next

Three things will decide whether GR00T 2 becomes infrastructure or a headline: the license terms when the weights and code arrive, independent replication of the demonstration-efficiency results, and whether the named industrial evaluators move from testing to deployment. Until then, the confirmed facts are strong enough on their own: a major open model, credible reported results, and a second front opened in the humanoid race within twenty-four hours of the first.

Sources and Further Reading

  • NVIDIA Blog, “Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation”, 9 November 2026: blogs.nvidia.com