Nvidia Releases GR00T 2, an Open Robotics Model That Learns From Human Demonstrations
Meet GR00T 2, our most advanced, open System One model for robotics research. It learns skills from human demonstrations and predicts how its actions will change the physical world. Those skills can carry across different robot bodies, reducing the need to teach each robot every task from scratch. Watch GR00T 2 put that research into motion. 馃У
Context
If skills learned from people can move across different robot bodies, labs will need far less robot specific training data.
Nvidia's official account posted on November 13 introducing GR00T 2, which it called its most advanced, open System One model for robotics research. The post says the model learns skills from human demonstrations, predicts how its actions will change the physical world, and can carry those skills across different robot bodies. Nvidia published a research page with its own test results for the model.
Deep dive
Nvidia's official account posted on November 13 introducing GR00T 2 as its most advanced, open System One model for robotics research. The post says the model learns skills from human demonstrations and predicts how its actions will change the physical world, and that those skills can carry across different robot bodies, reducing the need to teach each robot every task from scratch.
Nvidia's research page describes GR00T 2 as a foundation world action model that predicts future video and robot actions together, and says the models run on parallel grippers, dexterous hands and humanoids. According to that page, on 50 tasks never seen on the target robot, GR00T 2 succeeds 40% of the time with no demonstrations on that robot. Five demonstrations per task, which Nvidia says is under 3 minutes of robot data, raise that figure to 95% after a 30 minute fine tune on a single RTX 5090.
Nvidia also describes a faster variant, GR00T 2 Turbo, that it says can run on a Unitree G1 humanoid, with each prediction taking 124 ms on the robot's own Jetson Thor. Those figures come from Nvidia and have not been independently checked in the posts scanned.
What to expect next
Watch for Nvidia to publish weights, code or training recipes, and for other labs to try the model on their own robots. Independent tests of the official zero shot and few shot figures, and of skill transfer across robot bodies, would show whether the results hold outside Nvidia's setup. Follow up posts in the official thread may add demos or documentation links.