AI agents autonomously train robot fleet with 99% success

Researchers from Nvidia, Carnegie Mellon University, and the University of California, Berkeley have developed a system that enables artificial intelligence agents to autonomously train fleets of physical robots, achieving a 99 percent success rate on complex, real-world tasks. The framework, called ENPIRE, allows AI coding agents to design experiments, write and test software, and implement improvements on robot hardware without human intervention.

The system was tested on a fleet of eight dual-arm robots, which learned to perform precision tasks such as inserting graphics processing units into motherboard sockets, cutting zip ties, and sorting small pins into containers. According to the research team, the AI agents independently developed and compared training methods, retaining strategies that increased success rates. On some tasks, including pin insertion, the autonomous system achieved near-perfect reliability faster than comparable human-supervised training methods.

ENPIRE operates through a closed-loop system built around four core modules. An Environment module handles automatic resetting and verification of the robot's workspace after each attempt. A Policy Improvement...