Nvidia has unveiled GR00T 2, a new artificial intelligence model designed to teach humanoid robots delicate manual skills by learning from videos of people doing the same jobs. The company introduced the model on Monday 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside an updated simulation framework, Isaac Lab 3.0.

GR00T 2 is an open frontier world action model built on Nvidia's Cosmos platform. Rather than relying only on robot demonstrations, it was trained on diverse robot and human data, including 100,000 hours of Nvidia's open-source first-person human video. The idea is that a robot can watch how a person handles an object and then copy the movement, even for tasks it has never been shown on its own hardware.

How well does it perform?

In Nvidia's own tests, GR00T 2 completed 40 out of 50 evaluated tasks at least once without any task-specific demonstrations. With one demonstration per task, it reached an average success rate of 71 per cent across the 50 tasks, rising to 95 per cent after five demonstrations.

Those figures come with an important qualification. The technical paper behind the model notes that the tasks used for the five-demonstration test were selected from those the model could already learn well, so the average is high partly by design. The results are also from controlled tests, and Nvidia has not claimed the same success rates will hold in real factories and warehouses.

Nvidia says GR00T 2 ranks among the top three models across vision-language-action and world action model categories on benchmarks including RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with several leaderboard wins.

Training in simulation with Isaac Lab 3.0

Alongside the model, Nvidia released Isaac Lab 3.0, the latest version of its open-source robot learning framework. It lets researchers set up training environments quickly using software agents, switch between the PhysX and Newton physics solvers, and apply a new reinforcement post-training library to foundation models such as GR00T.

In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 running with the Newton physics engine delivered up to 2.5 times the training throughput of Isaac Lab 2.3 running with PhysX, according to Nvidia. The company also introduced Isaac Lab-Arena, an open-source framework for evaluating robot policies across thousands of simulated variations, with a developer preview of version 0.4 arriving at the conference.

Early users in factories

Several companies are already testing the technology. Foxconn is applying GR00T 2 to automate parts of its server manufacturing process, Siemens is evaluating it for industrial use, and Nexuni is testing it for laundry handling. Nvidia also named partners including Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington.

The model is described as open, but developers will have to wait for the code: Nvidia's research page says it will be available in December. For now, the announcement shows the direction of travel for robot AI, with human video and simulation doing much of the teaching before a robot ever touches a real object.