Nvidia unveils GR00T 2 robot model with world-action prediction
Nvidia says the model reached a 95% success rate on new tasks after five demonstrations per task and about 30 minutes of extra training.
Nvidia announced GR00T 2 and GR00T 2 Turbo, a robot AI model that predicts how the surrounding environment will change as a result of the robot's actions, at the Conference on Robot Learning (CoRL) in Austin, Texas, on November 9, 2026, according to aitimes.com.
GR00T 2 introduces a World Action Model (WAM) that predicts both robot motion and future changes in the environment. Nvidia said the model achieved an average success rate of 40% on 50 new tasks without additional demonstrations. With five demonstrations per task and about 30 minutes of additional training on a single RTX 5090, that figure rose to 95%, and a single added demonstration per task was enough to reach 71%.
The model was trained on more than 400,000 hours of first-person human task videos plus gripper demonstrations and teleoperation data. Learning from human demonstrations alone, GR00T 2 reached 65% on a funnel-and-filter-paper task and transferred a toy dump truck assembly skill to another robot arm with 60% success. Combined with a reasoning model, it assembled a toy airplane, tractor and car in sequence, with 97 to 98 successes out of 100 attempts for each.
Turbo runs on Jetson Thor inside a Unitree G1
GR00T 2 Turbo is a 2-billion-parameter model that runs directly on Nvidia's Jetson Thor inside the Unitree G1 humanoid. In an evaluation in unfamiliar environments, Turbo scored 63.3% against 40.0% for GR00T 1.7, and it succeeded in all 48 popcorn-serving demonstrations and 93% of table-tidying attempts.
Nvidia said Foxconn is using GR00T 2 to automate server manufacturing, Siemens is evaluating it for industrial sites, Nexuni is testing laundry handling and MicroAGI is using it to swap Ethernet cables with a robot arm.
Nvidia released its first general-purpose humanoid model, GR00T N1, in March 2025, and later followed it with the N1.5, N1.6 and N1.7 versions. On November 13, 2026, the company posted on X describing GR00T 2 as its most advanced open System One model for robotics research.
Quick answers
What is GR00T 2?
It is a Nvidia robot AI model that predicts both robot motion and future changes in the environment through a World Action Model.
How well does GR00T 2 perform on new tasks?
Nvidia reports a 40% average success rate on 50 new tasks without extra demonstrations, rising to 95% with five demonstrations per task and about 30 minutes of additional training.
What is GR00T 2 Turbo?
A 2-billion-parameter model that runs directly on Nvidia Jetson Thor inside the Unitree G1 humanoid.