# Nvidia opens GR00T 2 robot software in December

> GR00T 2 learns tasks from five human demonstrations and transfers them to another robotic body. The weights arrive in December, still without a license.
> By José Maldonado · Artificial Intelligence · 2026-11-18

A robot that learns to assemble a PC by watching a person do it, and then repeats the task with another body and other hands, stops being a brochure promise when someone publishes the hours of video it has used and the success rate of each test. Nvidia showed that arithmetic on November 9, 2026 at the CoRL conference in Austin when it presented [GR00T 2](https://research.nvidia.com/labs/gear/gr00t-2/), its action and world model for robotics.

The week closed with the promise that the weights and the Turbo variant will see the light in December. The figures accompanying the announcement frame the debate: with five human demonstrations of a new task, the model reaches an average 95% success rate, and without any demonstrations of its own it solves 40 of the 50 tasks it had never seen.

Those five demonstrations add up to less than three minutes of robot data, and the subsequent fine-tuning takes about 30 minutes on a single GeForce RTX 5090 card, the same one anyone buys to play games. The detail that changes the calculation for a lab is another one: skills travel between different robots, and until now, teaching a task required teleoperating each machine for hours and repeating the process for each model, so the cost was not in the software but in a human operator's time. If the skill transfers between bodies, that cost is paid once and amortized across every robot added to the workshop.

## What does a robot learn by watching videos of people?

GR00T 2 is not a language model with arms. It simultaneously predicts the video of what will happen in the following seconds and the movements that will produce that change, and then executes those movements. Nvidia calls that combination an action and world model, and built it on Cosmos, its simulation platform.

The training relies on three piles of data that the company details: 402,730 hours of video recorded by people with head-mounted cameras while they worked, 10,184 hours with manual grippers and 23,500 hours of remotely operated robots. That first block explains the sector's interest.

You don't need a robot to record yourself, just someone willing to wear a camera while doing their job. The list of tasks Nvidia published as a demonstration includes lighting matches, assembling toy trucks and preparing yogurt with banana, three examples that show variety rather than industrial difficulty.

The model covers arms with parallel grippers, dexterous hands and humanoids, and the company places it in the Isaac GR00T program alongside Isaac Lab 3.0, the simulation environment it released in October. The Turbo version deserves a separate paragraph because it is the one that can live inside the machine.

Each Turbo prediction takes 124 milliseconds on [the integrated Jetson Thor computer](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/). With it, a humanoid served popcorn in the 48 tests they ran and cleared a table in 93% of attempts, two short tasks that hide all the difficulty of grasping.

## The limits Nvidia itself acknowledges

The model does not process direct touch data, and that shows in tasks where force matters more than position. Researchers admit that it sometimes overlooks subtle errors, such as a screw left half-turned, and in a test of wiping up liquid with a towel underfoot it succeeded in fewer than half of attempts.

The second warning comes from outside and is worth keeping at hand. The specialist outlet TechRepublic noted that the 50 tasks in the experiment were chosen from among those that already exceeded 80% success after five demonstrations, so the initial 40% does not measure performance on any new real-world task.

The nuance of the prior selection does not invalidate the result, but it changes how to read it. A 40% on tasks filtered by the team itself is not the same as a 40% on tasks chosen by a client.

The third limit is pace. For long assemblies, Nvidia combines GR00T 2 with a language program that breaks the work into steps, and with that help the robots assembled toys in at least 97 out of every 100 attempts. The price of precision appears in trash sorting, where the robot waited on average more than seven seconds between decisions.

None of the three points invalidates the result, and all three together explain why a video of a humanoid serving popcorn is not yet an industrial contract. A robot that succeeds 93% of the time leaves 7% of broken hands or half-cleared tables, and that figure decides whether a plant accepts it alongside people.

## What can a lab test before December?

The downloadable part is already available and gets the fewest headlines. The Human-1M-100K dataset was published on [Hugging Face](https://huggingface.co/datasets/nvidia/PhysicalAI-Human-1M-100K) under a CC BY 4.0 license, which does allow commercial use. That is 100,000 hours of video from that larger corpus, available without asking anyone's permission.

Isaac Lab 3.0 has been openly available since October 20, and at CoRL the company also presented two evaluation benchmarks, DexBench and VABAR, to measure its own policies before touching a physical robot. That material allows work today, even without the model.

What you cannot do is download GR00T 2. The project page lists the code as available in December, with no exact date, no hardware requirements and no published license. The absence is not an administrative detail, because the license text decides what the announcement allows.

The weights of the previous generation, GR00T N1, arrived with an Nvidia license that restricts commercial use, and the Flora models were published under a non-commercial research license despite language similar to this announcement. Anyone planning a product will have to wait for the text, not the headline.

Meanwhile, a group with its own data can train on the published human video and compare its results with the open benchmarks, which is exactly the prior work that makes the model's arrival in December useful.

The company's context helps read the move. Nvidia opened [Nemotron for Science](https://observatorioblockchain.com/en/ai/nemotron-for-science-agentes-cientificos/) this same week with the same logic of giving away the software that makes its cards necessary, and robotics is where that strategy has the clearest return.

## The robotics to come will be played out in video

The shift GR00T 2 proposes is not in the model but in the data source. For a decade, robotics has depended on demonstrations made with the robot itself, a bottleneck that forced spending on hardware and engineer hours before knowing whether a task was possible.

Replacing much of that effort with video recorded on a camera changes the sector's unit of measure, which goes from robot hours to recording hours. Nvidia states that Foxconn applies the model in server manufacturing, that Siemens is evaluating it for industrial uses and that Nexuni is testing clothing handling in laundries with it.

These are names that place the announcement beyond the demonstration, although none of the published cases includes comparable results. Without that figure, what there is is a capability and not a result, and the difference between the two is what separates a video from a production line.

The useful question for the reader doing research is not whether the model is good, but whether their data is in a position to train it and whether they have a set of cases to verify it. The rest arrives in December, with a license that does not yet exist and that will decide whether robotics' next leap is built in the open or rented by token.
