Five demonstrations per task took NVIDIA’s new robot model from 40% to 95% average success in the company’s 50-task test. NVIDIA introduced GR00T 2 at the Conference on Robot Learning in Austin on November 9.

The model uses recent camera footage, an instruction and the robot’s current state to predict both its next moves and what the cameras should see after those moves. It then carries out a short sequence and looks again. In one research demo, a robot folds filter paper into a funnel and wets it with a wash bottle after seeing people demonstrate the task. NVIDIA reports that it succeeded in 65% of trials.

The 95% came after fine-tuning

Before tuning on demonstrations from the test robot, the model succeeded at least once on 40 of the 50 tasks. Across all attempts, it averaged 40%. One demonstration per task lifted the average to 71%. Five brought it to 95%, and the robot passed every trial on 37 of the 50 tasks. NVIDIA’s researchers say the five-demo run used “under 3 minutes of robot data” per task and a half-hour of fine-tuning on an RTX 5090.

Those figures describe NVIDIA’s lab evaluation. For a factory buyer, the number to ask for is how often an operator has to step in over a shift. NVIDIA says Foxconn is applying GR00T 2 to server manufacturing and Siemens is evaluating it. Its announcement gives no shift-level results from either company.

NVIDIA’s research page lists December for the code. Its technical report and the shipping-label demo are available now.