Nvidia introduced GR00T 2 on November 9 at the Conference on Robot Learning in Austin, Texas. It is a 2-billion-parameter diffusion transformer built on Cosmos-Predict2.5-2B, designed to teach robots new tasks from human demonstrations.
In a 50-task evaluation, success rose from 40% with no task-specific demonstrations to 71% after one and 95% after five. Each task needed under three minutes of robot data plus about 30 minutes of fine-tuning on a single RTX 5090.
Nvidia's own paper says the tasks were chosen from those that reached at least 80% success after five demonstrations, so the 95% figure is high partly by construction. Code is listed for release in December.
