AI agents test and improve robot code in virtual worlds before deployment

Researchers from UC Berkeley and NVIDIA have developed "agentic robotics," an approach where AI agents test and refine robot-control programs in virtual simulation environments to bypass the need for massive real-world training data. Using a framework called Graph-as-Policy and NVIDIA's Isaac simulation platform, AI models autonomously generate and debug code for robotic tasks. The method has already been applied in logistics by Ambi Robotics to increase parcel-sorting capacity. Open article →