NVIDIA's GR00T 2 learns skills from humans, not robots

Yuke Zhu announced GR00T 2, NVIDIA's new generation of what the team calls world-action models (models that predict how the scene will change and generate the actions to get there). One model runs on parallel grippers, dexterous hands and humanoids.

Yuke Zhu

@yukez

X

Much of what a robot learns need not come from robot data.

Yuke Zhu

@yukez

X

skills shown only by people, by hand or with wearable devices, transfer to robots that never practiced them

The part Zhu says he is most excited about is X-Adapt. Skills shown only by people, by hand or with wearable devices, transfer to robots that never practiced them. In the example he gives, a person makes breakfast wearing NVIDIA's in-house devices and the robot learns the task from those robot-free demos alone.

The data mix

GR00T 2 is trained on what Zhu calls the Data Pyramid. That is 400,000 plus hours of egocentric human video, 10,000 hours of UMI demos (data collected with handheld grippers rather than a robot), and 23,500 hours of robot teleoperation. Robot data is the smallest slice by a wide margin.

Agents can point at things

GR00T 2 can work alongside AI agents through a WAM API. A reasoning model plans the task and checks progress, GR00T 2 does the hands-on execution. Agents can also steer it by pointing, marking what to grasp and where to put it in the camera view, and the model handles the rest.

There is a faster variant, GR00T 2 Turbo, which NVIDIA says runs onboard a humanoid using Jetson Thor and drives the whole body as it walks, bends, reaches and carries. The listed tasks are serving popcorn, wiping up spills and clearing tables.

What is actually out

Code is not out yet. Zhu says GR00T 2 code is coming in December. What has shipped is data and hardware, 100,000 hours of NVIDIA's human video datasets on Hugging Face, plus open-source designs for the data collection hardware.

Everything above comes from NVIDIA's own announcement. No outside group has run GR00T 2 or GR00T 2 Turbo and reported results, and the post does not say how often the X-Adapt breakfast task succeeds or how long the onboard humanoid runs between failures. That arrives when the code does, or when somebody else gets their hands on it.