Teaching a robot a new job has meant hours of a person steering it through the task, over and over. NVIDIA's research team says its new model, GR00T 2, can pick up a new task from a single demonstration most of the time, and from five demonstrations almost every time.
NVIDIA introduced GR00T 2 on November 9, 2026, at the Conference on Robot Learning in Austin. NVIDIA calls it an open world action model built on its Cosmos platform. The model predicts what the scene will look like next and what the robot should do, both at once.
Key facts
- 71% average success on 50 tasks with one demo per task, and 95% with five.
- NVIDIA trained it on more than 400,000 hours of human video plus robot data.
- NVIDIA's research page says the code comes in December; it states no license yet.
What GR00T 2 does
NVIDIA tested GR00T 2 on 50 tasks that the target robot, a two armed machine with dexterous hands, had never practised. The model only saw those tasks in human video or in other robots' data.
| Training for the new task | Average success |
|---|---|
| No demonstrations | 40% of trials; 40 of 50 tasks succeeded at least once |
| 1 demonstration per task | 71% |
| 5 demonstrations per task | 95%; 37 of 50 tasks succeeded every time |
The five demonstrations add up to under 3 minutes of robot data. Each adaptation is a 30 minute fine tune on a single RTX 5090 graphics card, according to NVIDIA.
The pretraining mix includes 402,730 hours of first person human video, 10,184 hours of handheld gripper demonstrations and 23,500 hours of robot teleoperation. NVIDIA released 100,000 hours of that human video as an open dataset on Hugging Face.
Who can use it
NVIDIA named early users: Foxconn for server manufacturing, Siemens across industrial uses, Nexuni for laundry handling, and MicroAGI for swapping Ethernet cables with a robot arm. NVIDIA's research page lists the code as coming in December and does not yet give a license or a date for model weights. There is no price, since this is a research model.
What it means for a small business
You will not buy GR00T 2. It is a model that robot makers build on. The number that matters is the cost of teaching. If five short demos and half an hour on one graphics card can teach a robot a new task, robots for small runs and changing jobs get closer to making sense.
Think of a small packing operation, a lab or a kitchen that changes its tasks every few weeks. Robots have stayed out of those places because reprogramming took too long. GR00T 2 points at that cost coming down, though NVIDIA's own results show limits: most zero demo failures came from steps that needed more precision.
Our take
Based on the announcement, not on hands on testing: GR00T 2 is research news for robot builders, not a product you can order. Watch for two things. The December code release, and robot makers who sell systems that learn a task from a few demonstrations. When those reach your industry, ask vendors how many demos a new task needs and who records them.
Frequently asked questions
What is NVIDIA GR00T 2?
GR00T 2 is a world action model from NVIDIA research. It predicts future video and robot actions together, and it can adapt to a new task from a few demonstrations.
Is GR00T 2 open source?
NVIDIA calls it an open model. Its research page says code is coming in December and does not yet state a license.
How many demonstrations does GR00T 2 need?
In NVIDIA's tests, one demonstration per task gave 71% average success across 50 tasks, and five gave 95%.



