Jim Fan, NVIDIA's director of robotics, presented the model on Monday, the first day of the conference, at the close of an afternoon session.
According to Humanoids Daily, GR00T 2 predicts the next video frames at the same time as the robot's next actions, then carries out the actions it has just imagined.
NVIDIA reports an average success rate of 71% across 50 tasks with a single demonstration per task, and 95% with five demonstrations, after 30 minutes of tuning on a single RTX 5090 graphics card.
According to Jim Fan, the weights of GR00T 2 and of a lighter variant, GR00T 2 Turbo, will be released as open source in December, along with the training recipe for transfer between robot types.
NVIDIA names four companies testing the model: Foxconn, Siemens, Nexuni and MicroAGI, for factory work, laundry and cable handling.
Source
- Humanoids Daily · humanoidsdaily.com