NVIDIA GR00T 2 Teaches Robots New Skills From Human Video — 95% Success After Just 5 Demos
NVIDIA unveiled GR00T 2 on Monday 9 November 2026 at the Conference on Robot Learning in Austin, Texas, and the headline figure is striking: robots running the new model succeed at new tasks 95% of the time after seeing just five demonstrations. GR00T 2 is a robot foundation model — a single AI model that gives many different robots the ability to learn new physical skills, instead of each robot needing to be programmed task by task. It learns by watching first-person video of humans doing the job, and it was trained on 402,730 hours of that video.
What GR00T 2 is, and how it learns
GR00T 2 is an open frontier world action model built on NVIDIA Cosmos, and it is the successor to the GR00T N1 series. A world action model is a model that predicts what will happen next in the world and what action a robot should take, at the same time, inside a single diffusion transformer. It takes in camera views, a written instruction and the robot's current state, and it produces both future video frames and the robot's next movements together. That combination is what lets it plan a task and carry it out, rather than just recognising what it sees.
The training data is the real story here. The model learned from 402,730 hours of first-person human video, plus 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation, where a human operator drives the robot remotely to show it the task. NVIDIA is releasing 100,000 hours of that human data, called Human-1M-100K, under a CC BY 4.0 open licence, so other researchers can use it too.
The same skills transfer across different robot bodies, which NVIDIA calls cross-embodiment transfer. A skill learned from human video can run on a robot arm, on a dexterous five-fingered hand, or on the Unitree G1 humanoid, without being rebuilt for each machine. That matters because the robotics industry has dozens of different robot designs, and a model that only works on one of them is far less useful than one that works on all of them. For the wider context on where physical AI is heading, see our post on the NXP unified robotics platform physical AI, and the State of AI Report 2026 for the industry picture.
The benchmark numbers — and whose numbers they are
Every success rate in this section comes from NVIDIA's own tests, and that needs saying plainly. NVIDIA chose the 50 tasks, ran the tests, and published the results. No independent body has verified them yet, and a lab result is not the same as a factory floor. With that caveat, the pattern is clear. With zero demonstrations of a task, the model succeeds 40% of the time on average. With one demonstration per task, success rises to 71%. With five demonstrations and a 30-minute fine-tune on a single RTX 5090 graphics card, it reaches 95% average success across the same 50 tasks.
| Demonstrations per task | Average success rate | What it took |
|---|---|---|
| 0 demonstrations | 40% | No examples shown |
| 1 demonstration | 71% | One example per task |
| 5 demonstrations | 95% | 30-minute fine-tune on one RTX 5090 |
Zero demonstrations means the robot is guessing from its training, and 40% is already a useful starting point. One demonstration more than doubles the odds. Five demonstrations and half an hour of fine-tuning gets it to 95%, which is the level where a task starts to feel reliable. But 95% is still an average across 50 lab tasks, and the other 5% matters: a robot that fails one time in twenty will still break a production line, drop a product or need a human to step in. Factory reliability needs far more than 95%, and nobody should read this number as proof that robots are ready to run a warehouse unsupervised.
| Training data type | Hours |
|---|---|
| First-person human video | 402,730 |
| Handheld-gripper demonstrations | 10,184 |
| Robot teleoperation | 23,500 |
| Human video released openly (Human-1M-100K) | 100,000 |
NVIDIA also says GR00T 2 ranks in the top three across the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmark categories, again by its own measurement. The comparison with other world models is instructive: Google DeepMind Genie 4 generates interactive worlds from a single image, but it does not drive a physical robot. GR00T 2 is the other half of that idea, a model that turns a world understanding into real movement. Our Gemini Carbon vs Claude Opus 5.5 vs GPT-6 Astra comparison shows the same benchmark-table judging now reaching robotics.
GR00T 2 Turbo and Isaac Lab 3.0
Alongside the full model, NVIDIA launched GR00T 2 Turbo, a smaller variant built around a 2-billion-parameter transformer. It produces a prediction in 124 milliseconds on NVIDIA's Jetson Thor computer, which is fast enough for real-time use on the robot itself, without waiting for a server. A robot that waits for a distant computer before each movement cannot work safely around people. Turbo is the version meant to run on the machine, in the moment, and its speed is the feature that matters most for real deployment.
The same launch brought Isaac Lab 3.0, NVIDIA's open-source framework for training robots, which now adds the Newton physics engine and delivers up to 2.5 times the training throughput of Isaac Lab 2.3 on a dexterous-manipulation benchmark, according to NVIDIA. Faster training means researchers can test more ideas in the same time, and the open-source release means small labs can use the same tools as large ones. NVIDIA also launched two new open benchmarks for industrial robot work, giving the field shared tests to compare against.
Who is testing it — and what is not open yet
Several companies are already testing GR00T 2, and the list shows where the early interest lies. Nexuni is testing it for laundry handling, Foxconn for server-manufacturing automation, Siemens for industrial evaluation, and MicroAGI for swapping Ethernet cables with a YAM robot arm, post-trained with reinforcement learning. These are tests and evaluations, not deployments. Foxconn has not put GR00T 2 robots on a production line, and Siemens has not rolled it out across its factories. A test proves a task in a controlled setting; a deployment proves it all day, every day.
The openness question needs the same care. NVIDIA calls GR00T 2 an open model, but the model code is listed as available in December, and no licence or weights release date was announced at the launch. That means readers cannot download the model weights today, and nobody should promise that they can. What is genuinely open right now is the 100,000 hours of human video data under CC BY 4.0, and the Isaac Lab framework. The model itself is coming later, on a timetable NVIDIA has not fully spelled out.
The launch sits inside a bigger NVIDIA push into physical AI, and it is not the company's only news this week. Our post on the Nvidia Korea sovereign AI Rebellions SK Telecom partnership shows the same strategy from the other side: NVIDIA supplying the platform, and partners building on it. GR00T 2 is the robotics half of that platform, and its success will be judged by how many robots, in how many factories and homes, end up running on it.
Key Takeaways
- NVIDIA unveiled GR00T 2 on 9 November 2026 at CoRL in Austin: a robot foundation model that learns new skills from first-person human video.
- On NVIDIA's own tests across 50 tasks, success averages 40% with zero demonstrations, 71% with one, and 95% with five after a 30-minute fine-tune.
- Training used 402,730 hours of human video; 100,000 hours are released openly under CC BY 4.0.
- GR00T 2 Turbo, the smaller on-robot variant, predicts in 124 ms on Jetson Thor; Isaac Lab 3.0 adds the Newton physics engine.
- The model weights are not downloadable yet — code is listed for December, with no licence or release date announced — and Foxconn and Siemens are testing, not deploying.
Conclusion
GR00T 2 is the most convincing demonstration yet that robots can learn the way people do: by watching. Five demonstrations and half an hour of fine-tuning taking a robot to 95% success on a new task is a real step, and the 402,730 hours of human video behind it explain why. The caveats are just as real. The figures are NVIDIA's own, 95% is not factory reliability, the weights are not yet available, and the early testers are testing, not deploying. If the independent results match NVIDIA's claims, and the model arrives as openly as promised, the way robots are taught is about to change.
Comments
Post a Comment