NVIDIA's GR00T 2 learns mostly from human video, then needs five robot demos to hit 95%

NVIDIA introduced GR00T 2 on Nov. 9 at the Conference on Robot Learning in Austin, an open world-action model that predicts future video and robot actions together. Its research page lists what went into pretraining: 402,730 hours of egocentric video of people at work, 10,184 hours recorded with a handheld UMI gripper and 23,500 hours of robot teleoperation. Human video is about 92% of the hours; teleoperation is about 5%.

The result NVIDIA leads with shows where robot data still matters. On 50 tasks that the target robot, a two-armed machine with Sharpa hands, had never done but that appeared in human video or other robots' data, GR00T 2 succeeded in 40% of trials zero-shot. One demonstration per task raised that to 71%, and five, under three minutes of robot data, to 95% after a 30-minute fine-tune on a single RTX 5090, NVIDIA's researchers report. The broad skill comes from cheap video. The robot-specific part has shrunk to minutes per task, which hurts companies selling teleoperation by the hour more than it hurts rival model makers.

The misses are about precision, not about knowing what to do

NVIDIA's own reading of the zero-shot failures supports that split. GR00T 2 usually takes the right steps in the right order, the researchers write, and a single step falls short on precision. That is the gap a handful of demonstrations closes. With minimal post-training it seated GB300 memory modules with two 22-degree-of-freedom Sharpa hands in 85% of trials, and with a reasoning agent calling it as a tool it finished 292 of 300 toy assemblies.

Japan's AIRoA pushed the same idea further three days earlier. Its AIRoA-0 model, built on Physical Intelligence's π0.5 weights, was trained on 16,151 hours of handheld-gripper recordings with no robot teleoperation at all, and averaged 78% on ten tasks against 27% for the best of four other models trained on the same data, the association said on Nov. 6.

More video helps, but slowly

The strongest case against writing off teleoperation is in GR00T 2's own scaling test. NVIDIA pretrained four versions on none, 1%, 10% and 100% of the human video and measured them on 72 tasks. Real-robot task progress rose at every step, from 50% to 70%. A hundredfold increase in video bought 20 points, and the mix still holds 23,500 hours of teleoperation, so the robot hours have not gone away.

The claims will soon be testable outside NVIDIA. The research page says the code arrives in December, and 100,000 hours of the human video are already open on Hugging Face. Foxconn is applying GR00T 2 to server manufacturing and Siemens is evaluating it, according to NVIDIA. The number to watch is five: whether their engineers need five demonstrations per task on a real line, or fifty.