Nvidia announced a new AI model called GR00T 2. The system watches egocentric video of people doing tasks and then predicts both what the robot will see and how it should move. In lab tests a robot with five‑fingered hands could light a match, slice a banana or toss packages without any direct tele‑operation for those tasks.
GR00T 2 is described by Nvidia as a world‑action model (WAM). Unlike its predecessor GR00T 1, it does not need language commands paired with images. Instead it predicts the next video frame and the corresponding motor command together, using a diffusion transformer to denoise future video and actions.
How the model learns
The model was trained on a massive dataset that includes about 400,000 hours of head‑mounted video of people at work, 10,184 hours of demonstrations with a handheld UMI gripper, and 23,500 hours of robot tele‑operation. Nvidia says the human video makes up the bulk of the data and that scaling the amount of human video improves robot task success from 50 % to 70 % on real robots.
When a new robot needs to learn a task, a human can demonstrate the steps with data gloves or a finger exoskeleton. The system replaces the human arm in its video prediction with the robot’s arm and extracts the matching control commands. Nvidia reports that with just a few demonstrations, success rates rose from 40 % to 95 % after under three minutes of conventional training.
Performance and limits
In controlled experiments, a robot succeeded at hammering nails from a single human video and assembled toy cars in 97 % of trials when paired with a higher‑level language model such as GPT‑6 Astra. Without the language model, success dropped to 23 %. For bipedal robots, a slimmer version called GR00T 2 Turbo runs on the Jetson Thor onboard computer, taking 124 ms per prediction and handling tasks like scooping popcorn, clearing tables and wiping liquids, though liquid‑wiping succeeded in just under half of attempts.
Nvidia notes several hurdles: the model cannot yet use direct touch sensors, it sometimes misses subtle errors, and the language model adds a latency of about seven seconds per planning step, causing the robot to be idle for roughly a third of the time in trash‑sorting trials.
Why it matters
For manufacturers, the model could cut the time and cost of programming robot arms for new tasks, since a single human demo may be enough. However, the current need for external language models and the lack of touch sensing mean robots are not yet ready for fully reliable production lines.
