Nvidia and academic partners achieve 99% task success rate using autonomous AI coding agents to train robot fleets.
Nvidia, Carnegie Mellon University, and UC Berkeley researchers developed ENPIRE, a framework enabling AI coding agents to train robots independently. The system achieved a 99% success rate across tasks like pin insertion and GPU seating using agents such as OpenAI’s Codex and Anthropic’s Claude Code.
Testing with an eight-robot fleet reduced task mastery time by over 50% compared to a single robot, though token costs rose faster than time savings. The framework shifts AI-driven code testing from virtual to physical environments, where resets involve real-world robot movements.
The breakthrough highlights potential for scalable, autonomous robot training but underscores cost challenges as deployment expands.