Berkeley Researchers Release Open-Source Code for AI Agents to Build and Improve Robot Programs
Researchers at UC Berkeley, in collaboration with NVIDIA, have unveiled open-source code for AI agents designed to construct and refine robot control programs. This innovative approach, termed 'agentic robotics,' aims to alleviate the burdens of manual robot programming and the extensive data sets typically required for machine learning. The core of their method is the Graph-as-Policy (GaP) framework, which represents tasks as interconnected steps for perception, planning, and control, drawing from a modular library of existing robot skills. AI agents generate these connections and then utilize NVIDIA's Isaac simulation libraries to rehearse variations, identify failures, and modify programs before deployment to physical robots. This research, presented at the Conference on Robot Learning, includes an industrial-task benchmark to facilitate broader adoption by roboticists.