Nvidia GR00T 2 Teaches Robots Skills From Video: Dates, Demos & Limits
Nvidia's GR00T 2 model learns skills from human video, hitting 95% success after five demos — but code arrives in December and weights are not yet public.
Image: Wikimedia Commons / Coolcaesar
AUSTIN, Texas — NVIDIA introduced GR00T 2, a robot foundation model that learns new physical skills by watching videos of humans, at the Conference on Robot Learning in Austin on November 9, 2026. In NVIDIA's own tests, robots running the model completed new manipulation tasks with 95% average success after five demonstrations of each task, and completed 40 of 50 tasks at least once with none.
GR00T 2 is what NVIDIA calls a world action model. Built on the company's Cosmos family of world models, it predicts how a scene will change and which actions a robot should take, in a single model. The aim is to break the bottleneck that has slowed robot deployment: a robot learns from video of people at work instead of being programmed task by task.
Key Takeaways
- GR00T 2 averaged 95% success across 50 dexterous tasks after five demonstrations each, and 71% after one.
- Training used 402,730 hours of first-person human video, 10,184 hours of gripper demonstrations and 23,500 hours of robot teleoperation.
- NVIDIA is releasing 100,000 hours of that video as the Human-1M-100K dataset; model code is listed for December.
- Foxconn, Siemens and Nexuni are testing the model; none has put it into live deployment.
How Nvidia GR00T 2 Learns New Robot Skills From Human Video
The model learns through cross-embodiment transfer, a technique NVIDIA calls X-Adapt: it first learns the shape of a task from first-person human video, then transfers that understanding to a robot body. The 50 Dexter-50 tasks had no training data from the target robot at all.
| Demonstrations per task | Average success rate |
|---|---|
| 0 | 40% |
| 1 | 71% |
| 5 | 95% |
Success rates are NVIDIA's own figures, on tasks chosen among those the model learns at 80% success or better with five demonstrations.
A smaller variant, GR00T 2 Turbo, is built around a 2-billion-parameter transformer for humanoid robots, predicting its next action in 124 milliseconds on NVIDIA's Jetson Thor computer. NVIDIA's lead in AI hardware is under pressure from rival chips, including Google's latest TPU lineup.
When Will the GR00T 2 Code and Model Weights Be Available?
Not yet, and the distinction matters. NVIDIA's research page lists the model code as available in December, with no release date or licence announced for the model weights. What is available now is the data: 100,000 hours of the human video, released as Human-1M-100K under a Creative Commons CC BY 4.0 licence. Other labs are moving faster: Meta's Muse Ember agent model launched at 10 cents per million input tokens.
Which Companies Are Testing GR00T 2 and What Are They Using It For?
Foxconn is applying the model to server-manufacturing automation, Siemens is evaluating industrial uses, and Nexuni is testing laundry handling. All three are testing, not deploying: no rollout has been announced, and real-world reliability outside the lab is still unproven.
The 95% figure comes from NVIDIA's own tests, with no independent replication published. Demand for the computing behind models like this keeps climbing: CoreWeave raised its full-year revenue guidance after its order backlog reached $132.5 billion.
Frequently Asked Questions
What is Nvidia GR00T 2 and how does it work?
GR00T 2 is a robot foundation model unveiled by NVIDIA on November 9, 2026, at the Conference on Robot Learning. Built on NVIDIA's Cosmos world models, it predicts a scene's changes and the robot actions to take, learning new manipulation skills from videos of humans doing the task.
How many demonstrations does GR00T 2 need to learn a new task?
In NVIDIA's Dexter-50 evaluation, the model averaged 40% success with no demonstrations, 71% with one demonstration per task, and 95% with five demonstrations plus a 30-minute fine-tuning session. The 50 tasks were chosen from those it learns at 80% or better with five demonstrations.
Is GR00T 2 open source and when can developers download it?
Not yet. NVIDIA lists the model code as available in December 2026, with no release date or licence announced for the model weights. The 100,000-hour Human-1M-100K dataset is available now under a Creative Commons CC BY 4.0 licence, and the Isaac Lab 3.0 framework is open source.
Can GR00T 2 work on any robot?
That is the aim of its cross-embodiment transfer, which NVIDIA calls X-Adapt: skills learned from human video transfer to different robot bodies. The evaluation used a bimanual robot with two 22-degree-of-freedom hands.
The next date is December, when NVIDIA has said the GR00T 2 code will be available. The company has not said when the model weights will follow, or under what licence.
