Nvidia has unveiled GR00T 2, a new open robot foundation model that learns dexterous manipulation skills by watching first-person video of humans at work — a step toward robots that can pick up unfamiliar jobs without being shown each one in advance. The company introduced the model on November 9 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside Isaac Lab 3.0, the latest version of its open-source robot-learning framework, and a new open benchmark for testing robot behaviour.
GR00T 2 is a world action model built on Nvidia's Cosmos platform. Unlike earlier robot models trained mostly on robot-collected demonstrations, it is trained jointly on robot data and vast amounts of human video — including 100,000 hours of first-person human footage that Nvidia is releasing openly as the Human-1M-100K dataset, under a licence that permits commercial use. The idea is straightforward: humans already know how to fold laundry, plug in cables and handle tools, and a robot that has watched enough of that footage can transfer the skill to its own body.
Learning from watching, not just doing
The pretraining mix is enormous: more than 400,000 hours of first-person human video, over 10,000 hours of handheld-gripper demonstrations covering nearly 3,000 tasks, and 23,500 hours of robot teleoperation. A single pretrained checkpoint covers parallel-gripper arms, dexterous hands and full humanoids, so one model can serve very different machines.
In Nvidia's reported evaluations, the model completed 40 of 50 test tasks at least once without any task-specific demonstration — a measure of how well it generalises to jobs it has never been trained on. With a single demonstration per task, average success rose to 71%, and with five demonstrations it reached 95%. Nvidia also says GR00T 2 ranks among the top three models across several vision-language-action and world action benchmarks, including RoboCasa365, RoboLab, RoboDojo and MolmoSpaces. These are company-reported figures and have not been independently verified.
Simulation does the rehearsing
The second half of the announcement is about practice. Isaac Lab 3.0, generally available since October 20, lets researchers use AI agents to set up training environments with simulation-ready assets, switch between physics solvers, and apply reinforcement post-training to foundation models such as GR00T. A new AlphaNIST workflow applies large-scale simulated practice to 20 precision assembly tasks, including parts as small as four millimetres, while Isaac Lab-Arena supplies the evaluation layer.
Nvidia says the combination matters because delicate work — cable connections, small-part assembly — is where robots still fail most often, and simulation lets them rehearse thousands of times before touching real hardware.
Early testers, open code — with a wait
Early testing is already under way: Nexuni is trying GR00T 2 for robotic laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating industrial uses, while MicroAGI is using a reinforcement-learning post-trained version to swap Ethernet cables with a robotic arm. Nvidia describes all of these as tests and evaluations — no production rollout has been announced.
The model is presented as open, but there is a caveat: its research page says the code will be available in December, so researchers cannot yet download and run it themselves. The data release, the simulation tools and the benchmark are the parts available now. If the December code drop lands as promised, GR00T 2 could become one of the most widely built-on robot models of the year — and a test of whether watching humans really is the fastest way to teach a machine to use its hands.
