Researchers at NVIDIA have developed a system where teams of AI coding agents autonomously train robots to perform complex tasks such as inserting graphics cards and cutting zip ties.
The project relies on a framework called ENPIRE, created by NVIDIA's GEAR lab in collaboration with Carnegie Mellon University and the University of California, Berkeley. AI agents using models from OpenAI, Anthropic, and Moonshot AI tested and refined robot training policies over multiple cycles.
Experiments achieved a 99 percent success rate on tasks including moving blocks, organizing pins, and handling GPUs. Larger teams of up to eight agents completed training faster than smaller groups, though they consumed more computational resources.
Jim Fan, director of AI at NVIDIA, noted that the lab now self-improves overnight, with researchers reviewing results each morning. The team plans to open-source the framework for broader use.
The work builds on NVIDIA's recent robotics partnerships, including a May agreement with Unitree and discussions with Hyundai Motor Group in early June.