AI Agents Taught Robots to Install GPUs Overnight

Imagine leaving work for the day and returning in the morning to find your robots have learned new skills on their own. That's exactly what happened at Nvidia's robotics lab, where AI coding agents now run the overnight shift.

Researchers at Nvidia's GEAR lab, working with teams from Carnegie Mellon University and UC Berkeley, created a new system called ENPIRE. This framework lets AI agents take full control of robot training without human supervision.

The results surprised even the scientists. The AI agents developed training strategies that taught robotic arms to achieve a 99 percent success rate on tasks ranging from simple block pushing to delicate operations like inserting GPUs into motherboard sockets and cutting zip ties.

"We just read the reports in the morning," wrote Jim Fan, Nvidia's director of AI, describing how the system works tirelessly while humans rest. The team tested three different AI coding agents, including versions from OpenAI, Anthropic, and Moonshot AI, and each developed unique approaches to solving the training challenges.

The AI agents didn't just follow instructions. They analyzed failure logs, studied research papers, modified training code, and tested different strategies across multiple robots working simultaneously to find what worked best.

The Ripple Effect

The most exciting part might be what comes next. Nvidia plans to open source the entire ENPIRE system, meaning anyone could potentially set up their own self-running robot lab at home or in smaller research facilities.

In one particularly impressive test, the AI agents trained robots to organize pins faster than a cutting-edge human-supervised training method. Teams of eight AI agents working together achieved high success rates in just two hours, compared to nearly five hours for a single agent working alone.

The system isn't perfect yet. The robots often sat idle while the AI agents were busy debugging code or analyzing data. Larger teams of agents sometimes spent more time coordinating with each other than actually training robots.

The breakthrough comes as Nvidia pushes forward with ambitious robotics partnerships, including providing reference humanoid robots to research labs and working with Hyundai to scale up manufacturing of AI-powered robots. This self-improving training system could accelerate how quickly those robots learn useful skills.

What started as an experiment in autonomous research has opened a window into a future where robots might train themselves, freeing humans to focus on bigger questions while the machines handle the tedious work of trial and error.