NVIDIA Announces GR00T 2 Robot Model That Learns Tasks From Five Demonstrations
NVIDIA announced GR00T 2, a world action model for operating robots, trained on 402,730 hours of video of human movement, 10,184 hours of handheld gripper data, and 23,500 hours of robot teleoperation data.
1 source 1d ago
Reporting from 1 source: GIGAZINE.
Why it matters
NVIDIA is pitching the training data volume and the five-demonstration, 30-minute learning loop as the path from demonstration footage to working robot tasks, with the code release set for December 2026.
Key Facts
- NVIDIA announced GR00T 2, a world action model for operating robots.
- GR00T 2 was pretrained on 402,730 hours of video recording human movements.
- NVIDIA says a new task can be learned from five human demonstrations and executed in 30 minutes on a GeForce RTX 5090.
- Related code for GR00T 2 is scheduled to be released in December 2026.
- Part of the training data is published on Hugging Face as nvidia/PhysicalAI-Human-1M-100K.
NVIDIA says a new task can be learned from five human demonstrations and executed in 30 minutes on a GeForce RTX 5090, which can also run inference. Related code is scheduled for release in December 2026.
NVIDIA's GR00T 2 is trained on three separate data sets: 402,730 hours of video of human movement, 10,184 hours of handheld gripper operation, and 23,500 hours of human teleoperation of robots. The model handles whole-body motion, moving beyond hand movement alone.
NVIDIA describes tasks including preparing breakfast, assembling a PC, and lighting a match. In one demonstration the robot retrieves a cup from the floor, wipes spilled liquid with a towel underfoot, places the cup in a sink, and drops the towel in a collection box. GR00T 2 can also coordinate with other AI models, as in a Tower of Hanoi example run with GPT-6 Astra.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google