GR00T 2 learns robot skills from human video, and Nvidia calls it open

Nvidia has introduced GR00T 2, which it calls its most advanced open “System One” model for robotics research. According to Nvidia, it learns skills from human demonstrations and, in the company’s own tests, reached 95% average success on 50 tasks after five demonstrations each.

Nvidia’s post on X, published on November 13, describes the model as one that “learns skills from human demonstrations and predicts how its actions will change the physical world.” The model itself was first shown on November 9 at the Conference on Robot Learning (CoRL) in Austin, Texas, according to coverage of that announcement. The post is a fresh push for a release that is already several days old.

Numbers Nvidia reports for 50 unseen tasks

The figures below come from Nvidia and have not been independently verified. As summarized by The Next Web and Let’s Data Science, GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. With one demonstration per task, average success was 71%. With five, it was 95%.

Nvidia also says the model ranks among the top three in the vision-language-action and world-action-model categories on four benchmarks: RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with several leaderboard wins. One caveat appears in Nvidia’s own paper, as relayed by The Next Web: the 50 tasks were picked from ones the model could already learn well. The results therefore say less about arbitrary tasks than the headline numbers suggest.

The model is built on Nvidia Cosmos and trained on a mix of robot and human data, including 100,000 hours of first-person human video that Nvidia is releasing publicly. Nvidia says skills learned from people can transfer between robots with different bodies without robot demonstrations. An API and supporting tools let AI agents chain skills together and recover from errors.

Early testers named in the coverage are Foxconn, for server manufacturing, Siemens, which is evaluating it for industrial use, and Nexuni, which is testing laundry handling. MicroAGI has post-trained a reinforcement-learning version that swaps Ethernet cables with a robotic arm. According to The Next Web, none of the testers has committed to production use.

Predicting consequences is the hard part

Most robot policies map what a camera sees to a motor command. Nvidia describes GR00T 2 as a “world action model”, which also forecasts how the scene will change as a result of an action before the robot moves. That matters for soft or deformable objects such as cables and cloth, where a small error early in a motion compounds.

Learning from people has its own difficulty. A human hand and a robot gripper are shaped differently, so a video of a person folding a towel cannot be copied directly. Nvidia says its approach lets skills carry across robot bodies, and it released a companion open workflow, Video to Data, that turns human video into simulated robot demonstrations. Nvidia also reports that a 3D hand-motion reconstruction model trained with Noitom Robotics data passed 40,000 Hugging Face downloads in its first two weeks.

The labeling is worth noting. Nvidia’s post uses “System One”, a term from its earlier GR00T N1 design, where it meant the fast action-generating part of a robot model. The announcement coverage instead stresses the world-action framing. The sources reviewed do not explain how the two relate, so the architecture remains thinly documented.

Open, but terms and downloads are unclear

The model is described as open, and the training footage is described as public. However, the coverage reviewed gives no license, no download link and no date for model weights. One report said code is due in December, but that was not confirmed in the pages read here. Researchers who want to use it should check Nvidia’s research page and its Hugging Face account before planning around it.

GR00T 2 arrived with Isaac Lab 3.0, Nvidia’s open-source robot-learning framework. According to Nvidia, it adds a reinforcement post-training library for foundation models such as GR00T and integrates the open-source Newton physics engine to model cables, cloth and robotic hands. Nvidia reports up to 2.5x the training throughput of Isaac Lab 2.3 on one vision-based dexterous manipulation benchmark, and notes that results vary by workload.

The release fits a pattern of Nvidia supplying the models, simulators and data that robot makers build on. Nvidia’s leadership change this week has not altered that product cadence so far. Whether GR00T 2 performs outside Nvidia’s chosen benchmarks is the question that independent testing will have to answer.