Practice Makes Precise: NVIDIA and Partners Advance Robotics Simulation at CoRL

At the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA introduced GR00T 2, an open frontier world action model built on NVIDIA Cosmos. Trained on diverse robot and human data—including 100,000 hours of NVIDIA open-source human egocentric data—the model enables robots to learn dexterous manipulation skills across a wide range of physical tasks.

Key features of GR00T 2 include out-of-the-box performance, where the model completed 40 of 50 evaluated tasks without requiring task-specific demonstrations. It also demonstrates strong skill transfer, learning from human demonstrations and transferring skills between different robot bodies without additional robot demonstrations. Furthermore, GR00T 2 achieves high generalization with minimal data, reaching 71% average success across 50 tasks with just one demonstration per task, and improving to 95% with five demonstrations. The model also supports agent control, allowing AI agents to sequence skills and recover from errors using its application programming interface and supporting tools.

NVIDIA's advancements are part of a broader ecosystem of six finalists for the Best Paper Awards at CoRL, including papers like GRAIL and RoboTTT. The company is collaborating with partners such as Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Siemens, Skild AI, RLWRLD, UC Berkeley, Walden Robotics, and the University of Washington to disseminate these technologies to the robotics community. These innovations help researchers train robots on complex tasks, evaluate performance across changing conditions, and use failures to guide further learning—building the precision, adaptability, and autonomy needed for industrial deployment.

World action models represent the next level of robot autonomy, enabling machines to perceive their surroundings, anticipate environmental changes, and execute complex, multi-step tasks with high precision. GR00T 2 exemplifies this capability, ranking among the top three models across vision-language-action and world-action model categories at RoboCasa365 and RoboL benchmarks.