Robots have spent years getting better at moving through the world. The harder problem now is what they do when they arrive: aligning a gear, connecting a cable, or assembling a small component without crushing it. That shift from locomotion to delicate manipulation took center stage on Monday at the Conference on Robot Learning, known as CoRL, in Austin, Texas, where NVIDIA introduced GR00T 2, a new foundation model for robots.
NVIDIA calls GR00T 2 an open frontier world action model, built on its Cosmos family of world models. A world action model is meant to help a robot perceive its surroundings, anticipate how its actions will change the environment, and then carry out complex, multistep tasks with precision. GR00T 2 was trained on diverse robot and human data, including 100,000 hours of open-source human egocentric video, the first-person footage that shows tasks from the worker's own point of view. The point of that mix is dexterity: the model learns manipulation skills from human demonstrations and can transfer those skills between different robot bodies without needing fresh demonstrations collected on the robot itself.
The early numbers, all reported by NVIDIA, explain the interest. Out of the box, GR00T 2 completed 40 of 50 evaluated tasks at least once without any task-specific demonstrations. Given just one demonstration per task, its average success rate across the 50 tasks rose to 71 percent, and with five demonstrations it reached 95 percent. NVIDIA says the model currently ranks among the top three across all vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks, with multiple leaderboard wins. AI agents can also work with the model through an application programming interface, chaining skills together in sequence and recovering when something goes wrong.
Partners are already putting the model to work, although the operative words are testing and evaluating rather than deploying. Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases. MicroAGI is using a version post-trained through reinforcement learning to swap Ethernet cables with a YAM robotic arm, a deliberately fiddly task that shows why cables, which bend and slip, are such a useful test of dexterity.
Much of that training happens in simulation before a physical robot moves at all. Alongside the model, NVIDIA highlighted Isaac Lab 3.0, the latest release of its open-source robot-learning framework. It lets researchers use agents to set up training environments quickly, switch between the PhysX and Newton physics solvers, and apply a new reinforcement-learning post-training library to foundation models like GR00T. Tighter integration with Newton, an open-source physics engine, lets Isaac Lab model deformable objects such as cables and cloth, along with articulated robotic hands. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of the previous version with PhysX.
NVIDIA also added new ways to measure progress, including a developer preview of Isaac Lab-Arena 0.4 for running policies across thousands of evaluation variations, and two open benchmarks, DexBench and VABAR, designed to test dexterity and adaptability on industrial tasks. The results so far are the company's own, and benchmark success is not the same as reliability on a factory floor. Even so, the direction is clear: robot intelligence is moving from walking demos toward the fine, repetitive handwork that industry actually needs, and GR00T 2 is NVIDIA's bid to supply the brain for it.