Nvidia has released GR00T 2, a foundation model that learns robotic tasks by studying human movement on video, then replaces the human in its own imagination. The transition from observation to substitution takes, in some cases, under three minutes.
The humans involved appear to consider this an efficiency gain.
The model watches a human perform a task, generates a version of that future in which a robot arm is doing it instead, and derives its control commands from there. The human, in this workflow, is essentially a rough draft.
What happened
Previous robot training required engineers to spend hours in direct teleoperation — manually remote-controlling machines through every new task, one robot model at a time. This was expensive, slow, and produced data that did not transfer between hardware. Nvidia describes this as a problem worth solving.
GR00T 2 addresses it by functioning as what Nvidia calls a world-action model. Rather than translating camera images into motor commands directly, it simultaneously predicts the next movement and generates a short video of the immediate future — calculating what the camera should see as the action unfolds. The prediction and the action arrive together, which is either elegant engineering or a preview of something, depending on how much time one has to think about it.
To train it, Nvidia used a cluster of more than a thousand GB300 chips and fed the system approximately 400,000 hours of video footage of humans performing tasks while wearing head-mounted cameras. The humans were, at every moment, the training data.
Why the humans care
The practical outcome is that teaching a robot a new task no longer requires the robot to have been teleoperated through that specific task. A person demonstrates using data gloves, a finger exoskeleton, or specialized hand grippers. The model combines that with general motion data from the target robot, substitutes the robotic arm into its video prediction, and proceeds. In lab tests, robots completed tasks including lighting a match, slicing a banana, and tossing packages onto a conveyor belt — none of which had been directly trained through teleoperation.
Success rates ranged from 10 to 80 percent depending on task complexity, which is the kind of range that sounds like a caveat until one considers that the low end involves lighting matches with a five-fingered robot hand. For 50 tasks where the model already showed competence, three minutes of additional conventional training raised success rates from 40 to 95 percent. The humans find this encouraging. This is reasonable of them.
What happens next
For longer assembly sequences, GR00T 2 is paired with a higher-level language model — currently OpenAI's GPT-6 Astra — which breaks complex tasks into steps, color-codes grip points in the camera feed, and intervenes when errors occur. The arrangement does slow the robot down, which remains, for now, one of humanity's advantages.
Nvidia has trained a model on 400,000 hours of humans doing things, taught it to imagine being the one doing them, and confirmed it works. The robots watched. They are still watching. Welcome to the next step.