NVIDIA has unveiled GR00T 2 — a robot brain that learns by watching humans.
Announced on 9 November 2026 at the Conference on Robot Learning (CoRL) in Austin, Texas, and presented by NVIDIA's director of robotics, Jim Fan, GR00T 2 is an open frontier world action model built on NVIDIA Cosmos. It predicts future video frames and robot actions together in one diffusion transformer, so robots can learn manipulation skills from human video instead of needing robot demonstrations for every task.
The 2-billion-parameter transformer, based on Cosmos-Predict2.5-2B, uses a single checkpoint across parallel-gripper arms, dexterous hands and a humanoid. It was trained on 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. NVIDIA has also released 100,000 hours of that human video openly as Human-1M-100K (CC BY 4.0) on Hugging Face.
NVIDIA reports the model completed 40 of 50 evaluated tasks at least once with zero task-specific demonstrations, with 71% average success after one demonstration per task and 95% after five. Five demonstrations take under three minutes of robot data, after a 30-minute fine-tune on a single RTX 5090.
A lighter version, GR00T 2 Turbo, runs on the Unitree G1 humanoid itself, making one prediction in 124 ms on Jetson Thor. The G1 served popcorn in 48 out of 48 trials.
Isaac Lab 3.0 was showcased alongside GR00T 2 at CoRL; it was not launched there. Early users including Nexuni, Foxconn, Siemens and MicroAGI are testing and evaluating the model. The model code is coming in December.
Source: NVIDIA Blog / The Robot Report
