What researchers are saying

Fan shows GR00T 2 running long dexterous tasks on robot hardware

Jim Fan, robotics research lead at Nvidia

Fan posted uncut, real-time video of Nvidia’s GR00T 2 model controlling dexterous robot hands through tasks of more than 20 steps lasting over five minutes, using only on-robot compute.

Jim Fan, who leads robotics research at Nvidia, shared unedited video at normal speed of GR00T 2 driving real dexterous hands through long-horizon tasks. He said the runs covered more than 20 steps and lasted over five minutes, with all computation on the robot itself. He pointed to the model’s open weights, its ability to adapt skills across different bodies, and its integration with agent systems.

Nvidia’s research page describes GR00T 2 as a world-action model: a single diffusion transformer predicts future video and robot actions together, so it can learn physics from human video. The page reports that with two 22-joint Sharpa hands the model seats GB300 memory modules in 85% of trials, and that five demonstrations per new task raise success on 50 unseen tasks from 40% to 95% after a 30-minute fine-tune. Nvidia says the code will be available in December.

The step lengths and durations in the post are Fan’s own description of the videos.