Nvidia unveils GR00T 2 robot model trained partly on human video
The company reported 95% average success after five demonstrations per task, but said its tests used tasks the model could already learn well.
Nvidia introduced GR00T 2 at the Conference on Robot Learning in Austin, Texas, saying the open AI model helps robots acquire delicate manual skills using training that combines robot data with videos of people.
The training material includes 100,000 hours of first-person video, which Nvidia is making publicly available. The company says the model anticipates the visual effects of an action before performing it, and can transfer skills learned from people across robots built with different bodies.
According to Nvidia, the model completed 40 out of 50 tasks at least once without task-specific training. Average success reached 71% after a single demonstration and 95% after five demonstrations. However, Nvidia's research paper says the test tasks were selected from those the model was already able to learn well.
Foxconn is using GR00T 2 to help assemble AI servers, while Siemens is evaluating factory applications and Nexuni is testing laundry tasks. None of the three companies has announced a commitment to production use.
Nvidia separately released Isaac Lab 3.0, an open-source robot-training tool that uses simulation. The company reported training speeds up to 2.5 times faster in one test.
