NVIDIA unveiled GR00T 2, an open “world action model” for robot control, at the Conference on Robot Learning in Austin, the company said. Digital Today reported the same announcement, saying model weights and robot-specific training methods are due in December, while a 100,000-hour human egocentric dataset and related tools are available now.
Why it matters: NVIDIA said the model is built on Cosmos and trained on diverse robot and human data, including that 100,000-hour open dataset. It is designed to predict future video and robot actions together, and to run on robot arms, dexterous hands and humanoids.
In NVIDIA’s published evaluation, GR00T 2 completed 40 of 50 tested tasks at least once without task-specific demonstrations on the target robot, reached a 71% average success rate with one demonstration per task, and 95% with five, the company and Digital Today both reported. NVIDIA also said AI agents can call the model through an API to run skills in sequence and recover from errors. A lighter version, GR00T 2 Turbo, is meant to run on a Unitree G1 humanoid’s own Jetson Thor hardware.
NVIDIA said Foxconn is applying GR00T 2 to factory automation for server manufacturing, Siemens is evaluating it for industrial uses, and Nexuni is testing it for robotic laundry handling. Alongside the model, the company announced Isaac Lab 3.0, an open-source robot-learning update with tighter integration of the Newton physics engine for deformable objects such as cables and cloth.