NVIDIA GR00T 2 Teaches Robots by Watching Human Video
NVIDIA GR00T 2 learns robot skills from human video, hitting 95 percent success after five demonstrations. The catch: the code and weights do not arrive until December, and the numbers are NVIDIA's own.

NVIDIA introduced the GR00T 2 robot model on November 9, 2026, at the Conference on Robot Learning in Austin, Texas. It is a robot foundation model that learns physical skills by watching videos of people doing the work, instead of needing a robot to practice each task hundreds of times. That sounds like a small shift in training data. In practice it attacks the most expensive bottleneck in robotics: teaching a machine to do something new.
What the GR00T 2 robot model actually does
GR00T 2 is what NVIDIA calls a world action model. It takes in recent camera views, a written instruction, and the robot's current state, then predicts the next few frames of video together with the robot's next actions. The model imagines what the scene should look like a moment from now, then executes the movements that get it there. The core model has 2 billion parameters and is built on NVIDIA Cosmos, the company's platform of world models.
The training mix is the story. NVIDIA fed it 402,730 hours of first-person human video, 10,184 hours of demonstrations recorded with a handheld gripper across 2,897 tasks, and 23,500 hours of robot teleoperation spanning robot arms, dexterous hands, and the Unitree G1 humanoid. Human video is cheap and plentiful. Robot teleoperation is slow and costly. GR00T 2 is built on the bet that the first can substitute for much of the second.
The numbers, and the fine print
NVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstration on the target robot. With one demonstration per task, average success rose to 71 percent. With five demonstrations and a 30-minute fine-tune on a single RTX 5090, it reached 95 percent. A lighter variant, GR00T 2 Turbo, uses a smaller transformer and a more compact video representation for speed.
Read those figures the way they deserve to be read: as NVIDIA's own claims from its own evaluation suite. The company has not named the 50 tasks or published independent benchmarks, and no outside lab has reproduced the results yet. The 95 percent figure is the headline, but the 40 percent zero-shot figure is the more honest measure of where general-purpose robot learning stands today. It means the model still fails most new tasks outright until a human shows it the ropes.
Open, with an asterisk
NVIDIA calls GR00T 2 an open model, and that word needs a pause. The code arrives in December 2026, and the weights are promised for the same month along with the training recipe for transferring skills between robot bodies. As of the announcement, NVIDIA had published no license terms for the weights and no firm weights release date. What is available right now is the Human-1M-100K dataset, 100,000 hours of human video released on Hugging Face under CC BY 4.0, plus the research paper and project page.
That matters because the open label is doing real work in how this announcement is being received. The same month saw Kimi K3.1 Pro ship its full 2.8 trillion parameter weights openly, which shows what a real open release looks like, while Meta Muse Ember went the opposite direction with a tiny, cheap model priced for repetitive agent work. A model you can download and run is a different thing from a model you can read about. Until December, treat GR00T 2 as a research announcement with a delivery date, not as a tool you can use today.
Who is already testing it
NVIDIA named Foxconn, which is applying GR00T 2 to server-manufacturing automation, Siemens, which is evaluating industrial use cases, and Nexuni, which is testing robotic laundry handling. The announcement came alongside Isaac Lab 3.0, NVIDIA's open-source robot learning framework, the Newton physics engine, and new benchmarks called DexBench and VABAR. The package is deliberate: a model, a simulator, and a way to measure both.
The cross-body transfer approach, which NVIDIA calls X-Adapt, is the piece to watch. If a skill learned from human video can move between a factory arm, a dexterous hand, and a humanoid without starting over, the cost of adding a new robot to a production line drops sharply. If it cannot, GR00T 2 is an impressive demo that still needs a human expert for every new deployment.
What to watch next
December is the first real test. The open-weights field is moving fast, and Edge0-145B already claims to put a 145B model on a single GPU, so GR00T 2 will be judged against real downloadable models, not press releases. When the code and weights land, outside labs will try the same tasks on their own robots, and the gap between NVIDIA's numbers and everyone else's will show up fast. Watch for the license terms, too. A permissive license turns GR00T 2 into shared infrastructure for the whole field. A restrictive one keeps it as NVIDIA's showcase.
For now, the fair summary is this: robots that learn by watching people just got a lot more capable on paper, and the paper is the part you should keep in mind. The December release will tell us whether the capability survives contact with other people's robots.


