AI Agents Enable Robots to Rapidly Master New Skills

Researchers at the University of California, Berkeley have introduced a novel approach called 'Agentic Robotics' that leverages multiple AI agents to collaboratively write robot control programs. This method, utilizing a 'Graph-as-Policy' framework, decomposes complex tasks into modular skills which are tested and optimized in virtual environments using NVIDIA Isaac before deployment on physical robots. The system demonstrated its capability by enabling a robot to learn a wiping task from video data within an hour, automatically correcting for failures like pushing over objects based on real-world feedback. Currently, this technology is being applied by Ambi Robotics to optimize package sorting behaviors, resulting in each robot handling over 15,000 additional packages annually.