NVIDIA's AI agents taught robots to install GPUs into motherboards without any human help

NVIDIA has unveiled a framework that enables AI coding agents to train robots to perform high-precision physical tasks without human supervision. Developed alongside Carnegie Mellon University and UC Berkeley at NVIDIA's Generalist Embodied Agent Research lab, the system, called ENPIRE, closes the loop between writing robot training code, testing it on real hardware, and refining it until it learns the desired behavior.

The demo that caught everyone's attention shows a robot arm selecting a graphics card, passing it to a second arm, and carefully seating it into a PCIe slot on a motherboard. The test included a few other tasks, such as sorting metal pins into a box, cutting zip ties with real cutters, and the classic Push-T benchmark.

Across these contact-heavy tasks, the system achieved a 99% success rate under a pass@8 metric, which allows up to eight attempts per subtask with each retry informed by the previous failure.

3

ENPIRE works as a closed feedback loop with four parts. The environment module resets scenes and verifies results. Policy Improvement module writes and refines control code using reward signals, camera footage, and failure data. Rollout module runs physical trials on multiple robots simultaneously. An Evolution module then compares agent branches, keeping what works and dropping what does not.

Three AI coding agents were tested inside the framework: Codex on GPT-5.5, Claude Code on Opus 4.7, and Kimi Code on Kimi K2.6. Each developed its own algorithmic approaches, tested them on real hardware, and retained whatever improved the success rate. The agents share work through Git, which gives the whole system a practical research workflow feel rather than a polished demo.

3

A single agent took close to five hours to solve a task, whereas eight agents working in parallel cut that time to around two hours. However, larger teams also burn significantly more tokens, as agents spend more time reading each other's logs, summarizing branches, and coordinating, leaving robots idle while waiting for inference to complete.

NVIDIA plans to open-source ENPIRE, which would allow universities, startups, and hobbyists to run their own self-improving robot labs. Jim Fan, who co-leads the GEAR lab, described it as enabling autonomous research in the physical world for the first time. The next test will be whether it holds up outside NVIDIA's own controlled lab environment.