NVIDIA’s GR00T 2 Teaches Robots New Skills by Watching Humans — and It’s Going Open Source

November 11, 2026 Faizan Altaf

Teaching a robot a new job has always been slow, expensive work. Engineers usually have to program every movement by hand, or guide the robot through the same task hundreds of times before it gets it right. On 9 November 2026, NVIDIA announced a new AI model that could change that completely: GR00T 2, a robot brain that learns new skills largely by watching videos of humans doing the job.

NVIDIA introduced GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas, one of the biggest gatherings of robotics researchers in the world. The company calls it an “open frontier world action model” — and it plans to release the model openly in December, so researchers and companies around the world can build on it.

If it works as NVIDIA claims, the announcement could speed up the arrival of useful robots in factories, warehouses, hospitals and even homes. Instead of needing a team of engineers to teach each robot each task, a robot could pick up a new skill after watching a person do it just a handful of times.

What is GR00T 2?

GR00T 2 is the follow-up to NVIDIA’s first GR00T robot model, and it is built on NVIDIA Cosmos, the company’s family of so-called world models. A world model is an AI system that learns how the physical world behaves — how objects move, how they fall, how they can be picked up or pushed — and then uses that understanding to plan actions.

NVIDIA describes GR00T 2 as a “world action model” because it does two things at once. It predicts what will happen next in a scene, like a video of the future, and it also decides which actions the robot should take. In simple terms, the model imagines what it is about to do, and then does it.

The model is a 2-billion-parameter system — large enough to be capable, but small enough to run on a robot’s own computer. It was trained on a mix of robot data and human data, including 100,000 hours of open-source video showing the world from a person’s point of view.

How does it learn from watching humans?

The biggest problem in robotics has always been data. A chatbot can learn from trillions of words on the internet, but every hour of robot training data has to be collected by a real robot in the real world — slowly, and at great cost.

GR00T 2 attacks that problem in three ways.

First, it learns skills from human demonstrations and then transfers them to robot bodies. A person folding laundry or plugging in a cable shows the skill; the robot watches and works out how to do the same thing with its own arms and grippers, even though its body is different from a human’s. NVIDIA says this transfer works without needing demonstrations performed by the robot itself.

Second, it needs very little practice to learn something new. According to NVIDIA, GR00T 2 completed 40 out of 50 test tasks at least once with no task-specific demonstrations at all. With just one demonstration per task, it averaged 71% success across those 50 tasks. With five demonstrations, the average success rate rose to 95%.

Third, it can be controlled by AI agents. Software agents can use GR00T 2’s programming interface to chain skills together into longer jobs — pick up a part, fit it, check the result — and recover from mistakes along the way. That matters because real jobs are rarely a single movement. They are sequences of steps, and a robot that freezes at the first error is not much use on a factory floor.

NVIDIA is also releasing a new open tool called Video to Data, which turns ordinary human video into simulated robot demonstrations. The idea is to let researchers convert the world’s huge supply of video into training material for robots, without ever touching a real machine.

What the test results say — and what they don’t

NVIDIA reports that GR00T 2 ranks among the top three models in every vision-language-action and world action model category across four benchmarks: RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with several outright wins on the leaderboards.

Benchmarks are standardised tests that let researchers compare models fairly, so strong results across four of them is a good sign. But it is worth being careful about the headline numbers. The 71% and 95% figures come from NVIDIA’s own tests, on tasks the company chose, and the code is not public until December — so independent researchers have not yet had a chance to repeat the tests themselves.

That is normal for a new model announcement, and it is exactly why the open release matters. Once the weights and code are public, other labs can check the claims, find the weaknesses, and improve on the work. Until then, the fairest reading is that GR00T 2 looks very promising on NVIDIA’s own measurements, and the real test will come when other people try it.

Who is already using it?

NVIDIA says several companies are already testing GR00T 2 in real work:

Nexuni is testing it for robotic laundry handling — a fiddly job that involves picking up soft, floppy items that change shape, which is exactly the kind of task older robots struggled with.
Foxconn, the giant electronics manufacturer, is applying GR00T 2 to factory automation for server manufacturing.
Siemens is evaluating it across industrial uses.
MicroAGI is using GR00T 2, further trained with reinforcement learning, to swap Ethernet cables with a YAM robotic arm — a precise, delicate task that requires handling flexible cables.

NVIDIA is also working with partners including Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington.

These are tests and pilot projects, not full deployments. No one is claiming that GR00T 2-powered robots are ready to run a factory on their own tomorrow. But the list shows that the first customers are industrial companies with real production problems, not just university labs.

The announcement also included a smaller open model called EgoHand, which reconstructs 3D hand movements from video. It has already been downloaded more than 40,000 times on Hugging Face in the two weeks since its launch, according to NVIDIA — a sign of how hungry researchers are for tools that turn human video into robot training data.

The other half of the announcement: learning in simulation

A robot brain is only half the story. Before a robot can be trusted with a real task, it has to practise — and practising in the real world is slow and can break things. That is why NVIDIA also announced Isaac Lab 3.0, the latest version of its open-source robot-learning framework.

Isaac Lab is a simulation system: a virtual world where thousands of robots can practise at the same time, fail safely, and learn quickly. Version 3.0 lets researchers use AI agents to set up training environments fast, switch between two physics engines — NVIDIA PhysX and the open-source Newton engine — and apply a new reinforcement-learning library to foundation models like GR00T.

The physics matters more than it sounds. Tighter integration with Newton lets Isaac Lab simulate soft, bendable things such as cables and cloth, as well as complex jointed systems such as robotic hands with many moving parts. In a vision-based dexterous manipulation test, Isaac Lab 3.0 with Newton reached up to 2.5 times the training speed of the previous version with PhysX.

NVIDIA also showed how partners are using the simulation tools. Skild AI announced STEEL, a new physics solver built on Newton to speed up robot training, while Flexion, FieldAI and Dyna Robotics showed how Isaac Lab and Newton help train robots for factories, warehouses, hotels and food services.

One striking example came from a project called AlphaNIST, built in Isaac Lab. A single robot policy — one trained brain — learned all 20 tasks on the NIST Assembly Task Board, manipulating parts measured in millimetres, using only a yes-or-no signal for success. It reached more than 98% training success, compared with 81% for separate policies trained on one task each, while using roughly 39 times less computer time.

New tests to measure robot skill

Learning a skill is one challenge. Doing it reliably, in different places and conditions, is another. NVIDIA also announced two new open benchmarks designed to measure exactly that:

DexBench, built by RLWRLD with NVIDIA, tests dexterity in real-world workflows rather than on a tabletop. It has 15 tasks across four settings — retail, logistics, manufacturing and hospitality — and checks how well a robot copes when objects, positions and physics change.
VABAR, the Variational Automation Benchmark for Agentic Robotics, comes from UC Berkeley and NVIDIA. It tests robots across six industrial tasks, each with standard variations, difficulty levels and success measures.

Why this matters

Robots have been good at repetitive, fixed jobs for decades — a robot arm welding the same car part a thousand times a day, for example. They have been bad at jobs that change: picking up unfamiliar objects, handling soft materials, working in cluttered rooms, or adapting when something goes wrong.

That is the gap GR00T 2 is aimed at. If robots can learn from watching humans, and practise safely in simulation before touching the real world, the cost of teaching a robot a new job could fall sharply. Factories could reprogramme robots for new products in days instead of months. Warehouses could handle a wider range of goods. And further ahead, robots could help in homes, hospitals and care settings — places where no two days, and no two tasks, are ever quite the same. Robots are already moving into everyday life in other ways, from flying home cameras to warehouse machines, as we covered in our story on Amazon’s Ring drone camera and patrol robot.

There are also wider consequences. Robots that can do delicate, flexible work could change which jobs are automated first, and how quickly. Countries and companies that build the best robot brains could gain an edge in manufacturing and logistics. That is one reason NVIDIA, the world’s most valuable chip company, is investing so heavily in robotics software, not just the chips that power it.

As Ken Goldberg, a robotics professor at UC Berkeley who worked on two of the new projects, put it: “To me, agentic robotics is the most exciting technological development since the internet in 1995.”

That may turn out to be an overstatement — but it shows how seriously senior researchers are taking this shift.

What happens next?

The most important date is December 2026. That is when NVIDIA says the GR00T 2 model weights and code will be released, along with the training recipe for transferring skills between different robot bodies. A lighter version, GR00T 2 Turbo, is due at the same time.

Until then, researchers can study the architecture, the demonstrations and the published results, but they cannot download the model and try it for themselves. The December release will be the moment the wider robotics community can test NVIDIA’s claims — and start building on them.

Expect three things to watch: first, whether independent labs can reproduce the 71% and 95% success rates on their own robots; second, which companies move from testing to real deployment; and third, how quickly rival models appear. The race to build a general-purpose robot brain is now one of the most competitive fields in technology, with Google DeepMind, Tesla, Figure AI and others all working on similar ideas. For a closer look at how AI systems are being checked and verified, see our report on Google DeepMind’s AI proving Signal’s encryption secure in four weeks.

Frequently asked questions

What is NVIDIA GR00T 2?

GR00T 2 is an open robot foundation model from NVIDIA, announced on 9 November 2026 at the Conference on Robot Learning in Austin, Texas. It is a “world action model” built on NVIDIA Cosmos that learns dexterous manipulation skills — such as handling cables, folding laundry or assembling small parts — largely by watching videos of humans doing the same tasks.

How well does GR00T 2 perform?

According to NVIDIA’s own tests, GR00T 2 completed 40 out of 50 test tasks at least once with no task-specific demonstrations. With one demonstration per task, it averaged 71% success across the 50 tasks, rising to 95% with five demonstrations. These figures come from NVIDIA’s own testing and have not yet been independently repeated.

When will GR00T 2 be available?

NVIDIA says the model weights and code will be released in December 2026, along with the training recipe for transferring skills between different robot bodies. A lighter version called GR00T 2 Turbo is due at the same time. Some related tools, including the Isaac Lab 3.0 simulation framework and the EgoHand hand-tracking model, are already available.

What is Isaac Lab 3.0?

Isaac Lab 3.0 is NVIDIA’s open-source robot-learning framework. It provides a simulated world where thousands of robots can practise tasks in parallel before being deployed on real machines. The new version adds support for the Newton physics engine, which can simulate soft objects like cables and cloth, and reached up to 2.5 times the training speed of the previous version in NVIDIA’s tests.

Which companies are testing GR00T 2?

NVIDIA named Nexuni (robotic laundry handling), Foxconn (factory automation for server manufacturing) and Siemens (industrial uses) as early testers, along with MicroAGI, which is using the model for Ethernet cable swapping with a robotic arm. Partners also include Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Skild AI and RLWRLD.

Why does learning from human video matter?

Robot training data is scarce and expensive, because every hour of it must be collected by a real robot. Human video, by contrast, is abundant. If a robot can learn from watching people, it can borrow skills from the huge amount of human activity already on camera — making it far cheaper and faster to teach robots new jobs.

Conclusion

NVIDIA’s GR00T 2 is not a robot you can buy, and its headline numbers still need independent checking. But the idea behind it is powerful: robots that learn by watching humans, practise safely in simulation, and then transfer what they have learned to different bodies and jobs. Combined with the open release planned for December, faster simulation tools and new benchmarks to measure real skill, it is one of the clearest signs yet that general-purpose robots are moving from research demos towards real work. The next few months — when other labs finally get their hands on the model — will show whether the promise holds up.

Sources

This article is for general information and education only. It is not legal or financial advice.

Findivoza