NVIDIA expands simulation tools for precision robotics

NVIDIA introduced open robotics technologies at the Conference on Robot Learning in Austin, Texas, focused on teaching robots precise manipulation tasks such as connecting cables and assembling small components. The releases span world action models, simulation tools, evaluation benchmarks and AI agent workflows.

GR00T 2, an open world action model built on NVIDIA Cosmos, learns from robot and human data, including 100,000 hours of human egocentric recordings. It completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. Foxconn is applying the model to server manufacturing, while Siemens is evaluating industrial uses and Nexuni is testing laundry handling.

Isaac Lab 3.0 adds agent-assisted training setup, physics solver choices and reinforcement post-training tools. Integration with Newton enables simulation of deformable objects, including cables and cloth, alongside complex robotic hands. In a vision-based manipulation benchmark, the combination achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. These capabilities let researchers refine skills in parallel before transferring them to physical robots.

New benchmarks DexBench and VABAR test dexterity and adaptability under varying industrial conditions. UC Berkeley’s Graph-as-Policy framework uses agents to assemble individual skills into interpretable robot programs, test them in simulation and refine them after failures. Ambi Robotics, Bosch Research and Siemens are exploring the framework for industrial automation.