Show Nvidia’s GR00T 2 robot model a task five times and it succeeds 95% of the time

Nvidia‘s new robot model, GR00T 2, is built to learn a manual job from a few demonstrations instead of a long training campaign. Show it a task once and, in Nvidia’s own tests, it gets it right most of the time. Show it a few more times and it rarely misses. That matters because teaching a robot arm a new chore has usually meant steering it through the motions by remote control over and over, which is the slow and costly part of putting robots to work.

Nvidia calls GR00T 2 an open world action model. It is a kind of AI that looks at a scene, predicts how its movements will change it and then acts. It runs on Nvidia’s Cosmos world models and was trained on a mix of robot recordings and human footage. That footage includes a library of first-person video of people using their hands, which Nvidia has released as open source.

The numbers come from Nvidia’s evaluation of 50 manipulation tasks. With no demonstration specific to the task, GR00T 2 completed 40 of the 50 at least once. With a single demonstration per task, its average success rate was 71%, and with five it climbed to 95%. Nvidia says the model can watch a human do something and transfer the skill to a different robot body without any robot footage. AI agents can also call it through an API to chain skills together and recover when a step fails.

Some of those claimed abilities already have early testers. Taiwan’s Foxconn is applying GR00T 2 to factory automation for server manufacturing, Siemens is evaluating it across industrial jobs, and Nexuni is testing it for handling laundry. MicroAGI trained it further with reinforcement learning to swap Ethernet cables with a robotic arm.

Alongside the model, Nvidia shipped Isaac Lab 3.0, the open-source framework researchers use to train robots in simulation before risking real hardware. Its tighter link with Newton, an open physics engine, lets it simulate floppy materials such as cables and cloth, as well as robot hands with many joints. In one hand-manipulation benchmark, Nvidia measured up to 2.5 times the training throughput of the previous Isaac Lab 2.3. In practice that means more practice runs in the same amount of computing time.

There are limits to keep in mind. The success rates are Nvidia’s own measurements, and completing a task “at least once” is a long way from doing it all day on a factory line. GR00T 2’s top-three placings on robotics leaderboards such as RoboCasa365, RoboLab, RoboDojo and MolmoSpaces are also Nvidia’s claims, mostly in simulation. Nvidia describes the model as open, yet at the time of writing its weights had not appeared on the company’s GitHub or Hugging Face pages, where the newest downloadable version is still GR00T N1.7. Nor is there anything for consumers to buy. This is a toolkit for robot makers and labs, not a home robot.

For anyone outside a research lab, the change will show up slowly, in warehouse, factory and service robots that need less hand-holding before they are put to work. Because the tools are open source and published online, labs and startups anywhere can use them, from Nvidia’s partners in Korea such as RLWRLD to universities in Europe and Latin America.

Nvidia presented GR00T 2 and Isaac Lab 3.0 on November 9, 2026, at the Conference on Robot Learning in Austin, Texas. A developer preview of Isaac Lab-Arena 0.4, its tool for testing robot policies across thousands of variations, arrived at the same time, along with two new open benchmarks, DexBench and VABAR, built to measure how well robots handle industrial tasks. Nvidia has not given a date for the GR00T 2 weights to reach public download pages.

Tags: nvidia, robotics, humanoid robots, Foxconn, GR00T 2, Isaac Lab