Nvidia has unveiled GR00T 2, a robot “brain” that learns delicate hand skills partly by watching people. The company announced it on Monday 9 November at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside a new version of its Isaac Lab training software and two open tests for judging robot dexterity. The headline figures are striking, but they come from Nvidia’s own testing, and the model itself is not yet something you can download.

What Nvidia has announced

GR00T 2 is what Nvidia calls an open “world action model”. Instead of simply mapping a camera image to a movement, it predicts how the scene will look after an action, and chooses the action with that outcome in mind. It is built on Cosmos, Nvidia’s family of world-simulation models, and it is designed to run on simple two-fingered grippers, five-fingered robot hands and full humanoids.

The training diet is what stands out. According to Nvidia’s research page, the model learned from more than 400,000 hours of first-person video of people at work, around 10,000 hours of demonstrations using a handheld gripper, and about 23,500 hours of robot teleoperation. Nvidia is releasing 100,000 hours of the human footage openly, so other researchers can use it too.

Alongside the model, Nvidia announced Isaac Lab 3.0, the latest release of its open-source framework for training robots in simulation. It also introduced two open benchmarks, DexBench and VABAR, and a workflow called Video to Data that turns ordinary human video into simulated robot demonstrations.

The numbers, and the small print

Nvidia tested GR00T 2 on 50 tasks that appeared in other robots’ data or in human video, but never in the data of the robot doing the test, which had two five-fingered hands. With no task-specific demonstrations at all, the model completed 40 of the 50 tasks at least once, and succeeded about 40% of the time on average. Given a single demonstration of each task, average success rose to 71%. With five demonstrations, it reached 95%.

Those five demonstrations amounted to under three minutes of robot data per task, followed by a fine-tuning run of around 30 minutes on a single RTX 5090 graphics card.

There are caveats. Every figure is Nvidia’s own, and nothing has been independently tested. The Next Web reports that Nvidia’s own paper notes the 50 tasks were chosen from ones the model could already learn well, so the 95% average partly reflects how the tasks were picked. Automatica Press likewise notes that the figures are Nvidia’s own and have not been independently tested.

Learning from people, not just robots

The most interesting result is not the headline average but what Nvidia calls cross-adaptation. In one test, a robot with two five-fingered hands lined a laboratory funnel with filter paper and wetted it, a task it had only ever seen people do, with no robot demonstration. It succeeded in 65% of trials. In another, a robot arm assembled a toy dump truck end to end in 60% of trials after seeing the job done only with a handheld gripper, a different tool from its own.

If that holds up, it matters because human video is far cheaper and more plentiful than robot data. Collecting teleoperated demonstrations needs a robot, an operator and time, whereas people can film themselves doing chores or factory tasks with a head-mounted camera. Startups are already paying people to do precisely that, and a model that can turn such footage into robot skills would ease one of the field’s biggest bottlenecks.

Agents in charge, humanoids on board

Nvidia also describes GR00T 2 as something an AI agent can drive. A reasoning model can call the robot model through an interface, string skills together and recover when something goes wrong. In Nvidia’s tests on a robot arm, that set-up completed 97 or more of 100 back-to-back toy assemblies for each of three toys, compared with 71 to 84 for the same robot policy working alone.

A smaller variant, GR00T 2 Turbo, is light enough to run on the robot itself. Nvidia says it runs on a Unitree G1 humanoid’s onboard Jetson Thor computer, taking 124 milliseconds per prediction, and in tests the robot scooped and served popcorn in all 48 trials. Ken Goldberg, a UC Berkeley roboticist who worked with Nvidia on two of the projects, called agentic robotics “the most exciting technological development since the internet in 1995”.

Training in simulation gets faster

Isaac Lab 3.0 is the less glamorous half of the announcement, but it is the part aimed squarely at developers. It integrates more tightly with Newton, an open-source physics engine, so simulated robots can now handle soft things such as cables and cloth, and hands with many joints. Nvidia says that on one dexterous-manipulation benchmark it trains up to 2.5 times faster than Isaac Lab 2.3 running on the older PhysX engine. It also adds a reinforcement-learning library for fine-tuning foundation models such as GR00T.

Partners are already building on it. Skild AI announced a new physics solver for Newton, and Flexion, FieldAI and Dyna Robotics described using the tools to train their own robots.

Who is using it, and when you can get it

Nvidia says Foxconn is applying GR00T 2 to factory automation for server manufacturing, Siemens is evaluating it for industrial use, Nexuni is testing it for handling laundry, and MicroAGI is using a version trained further with reinforcement learning to swap Ethernet cables. The Next Web notes that none of them has said it will use the model in production yet.

Availability is also still vague. Nvidia calls GR00T 2 open, but its research page says the code will be available in December, and Automatica Press reports that an Nvidia page gives “by the end of the year”. The company gave no pricing in the announcement.

Why it matters

For years the humanoid race has been framed as a contest of hardware: who can build the most agile body at the lowest price. GR00T 2 is a reminder that the bottleneck is increasingly the software and the data. If a robot can pick up a new job from a few minutes of demonstration, or from watching a person, the economics of deploying robots in small factories, labs and warehouses could change, and so could who gets to build them.

That is a theme the UK has an interest in, with start-ups such as Machenta betting on teaching robots from a single demonstration. It is worth remembering how often impressive lab results take years to turn into dependable products. The next things to watch are whether independent researchers can reproduce Nvidia’s figures once the code is out in December, and whether Foxconn or Siemens move from testing to putting GR00T 2 on a production line.

Keep exploring