NVIDIA Research has published GR00T 2, a family of robot models that picture the next few seconds of video while they work out the motor commands to make it happen. NVIDIA calls them world-action models. The bet is that a robot able to imagine the result of a movement can learn from footage of people, and footage of people is far cheaper to collect than robot demonstrations. Every number below comes from NVIDIA Research’s own project page, tested on NVIDIA’s own robots and tasks.
The training mix shows where the weight sits. It combines 402,730 hours of first-person footage of workers, 10,184 hours recorded with a handheld gripper (2,897 different tasks), plus 23,500 hours of humans remote-controlling robots. Human video dwarfs the rest. To see whether it earns its place, NVIDIA trained four models with none, 1%, 10%, and all of it, then ran 72 tasks on its Dexter robot. Task progress on the real machine climbed from 50% to 70%, and in generated video rollouts from 64% to 77%.
The most striking claim is that a robot can pick up a task it never practiced. Shown only human demonstrations, a robot with two hands and five fingers each lined a chemistry funnel with filter paper, then wet it, in 65% of trials. Shown only a handheld gripper, a different machine called YAM assembled a toy dump truck in 60% of trials, on a job that runs up to four minutes.
Fast adaptation is where the figures get large. On 50 tasks that existed only in other robots’ logs or on human footage, GR00T 2 succeeded 40% of the time with no robot-specific training. Five demonstrations per task, under three minutes of robot data, pushed it to 95%, after half an hour of extra training on one RTX 5090 graphics card. NVIDIA says most early misses came from one step lacking precision, with the order of steps right.
A second layer puts a separate reasoning model in charge of planning while GR00T 2 handles the hands. With that setup, YAM assembled a toy plane, a toy tractor, and a Hot Wheels car, succeeding 98, 97, and 97 times out of 100 attempts. The same trained model working alone managed between 71 and 84. When the planner also drew grasp and drop points onto the camera view, success on short what-and-where instructions jumped to 92% from a 40% baseline on YAM.
The limits sit in the same tables. On the Dexter hands, those marks got more items into the bin but seldom finished the whole job, which NVIDIA names as the system’s current limit. The compact version, GR00T 2 Turbo, has 2 billion parameters and runs on the Unitree G1 humanoid’s own Jetson Thor chip at 124 milliseconds per prediction. It served popcorn successfully every time across 48 trials and cleared a table in 93%, yet it trails the full model on most of the other tasks NVIDIA compared.
Nobody outside NVIDIA has reproduced any of this. The tasks, the robots, and the scoring are all NVIDIA’s, and a 40% zero-shot rate still means most attempts at an unseen task fail. One check is more reassuring: across 14 policies, simulated scores tracked real-world scores with a Pearson correlation of 0.92, which suggests cheap simulation can stand in for some robot time.
The practical opening is on the data side. NVIDIA has put 100,000 hours of the human video on Hugging Face as Human-1M-100K, so a robotics team can test whether first-person footage lifts its own machine before paying for another round of teleoperation.
Reported by NVIDIA Research on its GR00T 2 project page; the page carries no publication date.
