Nvidia introduced the GR00T 2 model, which can learn tasks from videos of people or other robots and predict how the environment will change after actions. In tests on 50 tasks, the model succeeded in 40% of attempts without prior demonstrations, 71% with one demonstration per task, and 95% with five demonstrations.
Each task required less than three minutes of demonstration data and about 30 minutes of fine‑tuning on a single Nvidia RTX 5090 GPU. The system uses a World‑Action Model architecture that receives recent camera frames, a text instruction and robot state, then generates future video frames and corresponding motions.