GR00T 2 is NVIDIA's open world-action model for teams building manipulation and humanoid policies. It reads recent camera views, a language instruction, and the robot's state, then predicts the next camera frames together with the actions that produce them. The same family covers parallel grippers, dexterous hands, and humanoids. GEAR introduced it on November 9, 2026, at the Conference on Robot Learning in Austin. Jim Fan, NVIDIA's director of robotics and co-lead of GEAR, announced it as an open System 1 model for long-horizon work, including dexterous manipulation and whole-body control chained across more than 20 steps and more than five minutes. The onboard results use GR00T 2 Turbo. On a Unitree G1, Turbo runs on the robot's Jetson Thor at 124 milliseconds per prediction and sends 1.6-second motion chunks to the SONIC controller, which executes them at 50 Hz. The reasoning model remains off the robot. On 50 tasks the target robot had never practiced, NVIDIA reports 40 percent success with no robot demonstrations, 71 percent after one, and 95 percent after five, from under three minutes of robot data and a 30-minute fine-tune on one RTX 5090. Weights, the cross-embodiment recipes, and code are scheduled for December. Available now are 100,000 hours of NVIDIA's open human egocentric video, hardware designs for handheld grippers and wearable hands, and the agent, simulation, and annotation tools. Foxconn is applying the model to server manufacturing, Siemens is evaluating industrial uses, and Nexuni is testing laundry handling. The bet for operators is that human video plus a handful of robot demonstrations can replace a long teleoperation campaign for each new task.