NVIDIA Unveils GR00T 2: Robots Learn Tasks From Human Videos With 95% Success
NVIDIA unveiled GR00T 2, a next-generation open-world action model for humanoid robots, at the Conference on Robot Learning (CoRL 2026) in Austin, Texas, on November 9, 2026. The model marks a significant step toward robots learning dexterous physical tasks by watching people.
Built on NVIDIA’s Cosmos platform, GR00T 2 was trained on a dataset that includes 100,000 hours of open-source, first-person human video. It can transfer skills from human demonstrations to robot movements without requiring robot-specific training data.
In NVIDIA’s internal benchmarks, GR00T 2 achieved the following results:
- Completed 40 of 50 evaluated tasks with no task-specific demonstrations.
- Achieved a 71% average success rate with one demonstration per task.
- Reached a 95% success rate with just five demonstrations.
Alongside GR00T 2, NVIDIA introduced Isaac Lab 3.0, an upgraded open-source framework for robot-learning simulation. Industry partners—including Foxconn, Siemens, Nexuni, and MicroAGI—are already testing the model in real-world factory, laundry, and cable-handling applications.
NVIDIA’s director of robotics, Jim Fan, presented the 2-billion-parameter model at CoRL, describing the initiative as a step toward making robotics accessible to more developers.
Sources: Unite.AI | Blockchain.News