NVIDIA GR00T 2 masters new robot tasks from five demonstrations with 95% success
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Teaching a robot a new job may now require only a few short examples and half an hour of fine-tuning on a consumer graphics card: different robots are beginning to learn skills from one another and directly from people.
NVIDIA has introduced the open GR00T 2 model at CoRL, which predicts future video and robot actions together. In the company's own tests on 50 tasks the target robot had not previously performed, the model achieved an average success rate of 40% without demonstrations and completed 40 tasks at least once. One demonstration raised the average result to 71%, and five to 95%; the five recordings took less than three minutes, while LoRA fine-tuning took 30 minutes on a single RTX 5090. The model transfers skills from videos of people, the handheld UMI gripper and robots to other robot bodies. Pretraining used 402,730 hours of egocentric video, 10,184 hours of UMI demonstrations and 23,500 hours of robot teleoperation. Paired with a reasoning agent that breaks work into steps and recovers from failures, GR00T 2 assembled three toys in 292 of 300 continuous trials. The lighter GR00T 2 Turbo runs directly on the Unitree G1 humanoid's Jetson Thor: a prediction takes 124 ms, and clearing a table of assorted objects succeeded in 93.3% of trials. All figures are NVIDIA's results on its own task sets.
Sources 2