Nvidia GR00T 2: Robot model that learns tasks from a few examples

What's new

Nvidia has unveiled Isaac GR00T 2, a robot model meant to be used for many different tasks and robot types. It builds on the DreamZero research project and on what is known as a world action model. That means the model both predicts what the world will look like a moment from now and which actions the robot should perform. It does this in a single pass.

The model has 2 billion parameters (the adjustable knobs an AI model learns with) and is built on Nvidia Cosmos. It was trained on 402,730 hours of first-person video of humans, 10,184 hours of demonstrations with a handheld gripper spread across 2,897 tasks, and 23,500 hours of robot teleoperation, where a human controls the robot. It has been tested on the YAM and Franka robot arms, hands from Sharpa and Wuji, and the Unitree G1 humanoid, among others.

What's smart

The point is that the robot should learn quickly. According to Nvidia, the model completes 40 percent of 50 tasks without having seen a single demonstration (zero-shot). With one demonstration per task, the average is 71 percent, and with five demonstrations, equivalent to under three minutes of robot data, it is 95 percent. The 95 percent applies to tasks selected because they can be learned with five demonstrations (at least 80 percent success), so the figure is high partly by design. All figures are Nvidia's own measurements.

Adapting to a new task requires about 30 minutes of fine-tuning on a single RTX 5090 graphics card. Nvidia also cites 85 percent success at handling GB300 memory modules and 60 percent on a four-minute assembly of a toy truck. The GR00T 2 Turbo variant is the same size but uses four times fewer spatial tokens (small pieces of visual information). It runs on Jetson AGX Thor at 124 milliseconds per prediction.

Cheaper or better?

The price has not been disclosed. According to Nvidia, robots succeed at new tasks in new environments more than twice as often with GR00T 2 as with leading vision-language-action models, and the model ranks No. 1 on MolmoSpaces and RoboArena for generalist robot models. That is Nvidia's own claim, and the competitors are not named.

The model weights (the fully trained model itself) are expected to be released in December 2026. Already available now are the SimFoundry pipeline, the Flora annotation models, the Human-1M-100K dataset with 100,000 hours of data, and hardware designs. The license for the model itself has not yet been published, and Flora is released under a non-commercial research license. The sources do not say what availability looks like in Denmark and the EU. Nvidia researcher Jim Fan says robotics is no longer an elite sport and that you should own your System 1, unbound and free.

What it's good for

The model is aimed at robots that need to perform physical tasks, and where it is expensive to show the robot everything from scratch. According to the article, Foxconn, Siemens, Nexuni and MicroAGI are testing it in production, laundry and cable handling.

For ordinary users, there is nothing to buy or download yet. For companies with robots, the short setup times could be interesting once the weights and the license are in place.