NVIDIA's open-source GR00T 2 learns fine dexterous manipulation from human videos

On November 9, 2026, NVIDIA presented advances in robot simulation and manipulation at CoRL 2026, the Conference on Robot Learning in Austin, Texas. An NVIDIA blog post framed the shift from locomotion to manipulation: simulation, long the proving ground for robots learning to walk, roll and hop, is now helping researchers teach robots tasks requiring a finer touch, such as aligning gears, connecting cables and assembling small components. At the same event, NVIDIA introduced GR00T 2, Isaac Lab 3.0, and open benchmarks for robot training.

GR00T 2 is described as an open frontier world action model built on NVIDIA Cosmos. Accounts differ on its training data and reported performance. A DEV Community post carrying the NVIDIA blog text says the training combines robot and human data, including 100,000 hours of NVIDIA's open source human egocentric data, and that AI agents can transfer skills learned from human demonstrations between different robot bodies without robot demonstrations. The Next Web reports that the model was partly trained on 100,000 hours of human video to help robots learn delicate hand skills, reached 95% average success with five demonstrations per task in Nvidia's tests, and is being tested by Foxconn, Siemens and Nexuni. The Decoder reports a much larger training corpus of about 400,000 hours of video, with human hand movements extracted from video and transferred to robotic arms without laborious teleoperation, and success rates that vary widely from 10% to 80% depending on the task. Those figures have not been reconciled, and the available reporting stops at the launch, with the discrepancies between the tested success rates left unresolved.