If you have ever fought with a PCIe retention clip while installing a graphics card, you will appreciate just how delicate the process can be. NVIDIA researchers have now demonstrated robots that can learn to perform exactly that task, inserting GPUs into motherboards without direct human guidance.
The work comes from NVIDIA’s Generalist Embodied Agent Research, or GEAR, lab in collaboration with researchers from Carnegie Mellon University and the University of California, Berkeley. Their framework, called ENPIRE, allows AI coding agents to autonomously train robots through trial and error rather than relying on painstaking hand-written instructions.
Videos released by NVIDIA show robotic arms successfully aligning and inserting graphics cards into PCIe slots, sorting tiny metal pins, and cutting zip ties. These are not abstract software benchmarks, but physical tasks that require precise positioning, careful force control, and repeated refinement after failure.
ENPIRE effectively turns coding agents into autonomous robotics researchers. Instead of engineers manually designing every training routine, the AI agents generate code, evaluate results, debate possible improvements, consult technical papers, and continuously modify their control policies.
NVIDIA researcher Jim Fan described the system as “AutoResearch in the physical world,” where AI agents are given access to robots, GPUs, and a token budget, then assigned a task to solve. The idea is that the agent does not merely follow instructions, but actively searches for better ways to make the robot succeed.
The ENPIRE name represents four components: Environment, Policy Improvement, Rollout, and Evolution. Together they create a feedback loop in which robots repeatedly experiment, analyze failures, and improve their behavior. Multiple robots can also work in parallel, allowing the system to gather more experience much faster than a single machine could.
That parallel approach is one of the more important parts of the research. Adding more robots means more real-world attempts, more failure data, and faster policy optimization. In some of the demonstrations, eight robots were used simultaneously.
Researchers also compared different coding models, including OpenAI’s Codex, Anthropic’s Claude Code, and Moonshot AI’s Kimi Code, to see how well each could help train robotic systems. The GPU installation demo has naturally grabbed the attention of PC enthusiasts, but the real story is bigger than a robot slotting a graphics card into a motherboard.
Industrial robots have traditionally required extensive programming, tuning, and specialist supervision before they can take on new physical tasks. ENPIRE is beginning to lay the groundwork for a different model, where AI agents help robots adapt faster, with less manual engineering effort and fewer human-led training sessions.
The project also fits NVIDIA’s broader push into what CEO Jensen Huang has called “physical AI,” combining foundation models, simulation environments, robotics platforms, and accelerated computing into systems that can interact with the real world.
Human PC builders probably do not need to worry just yet. Watching NVIDIA’s robots install graphics cards does bring back memories of TigerDirect’s old PC Race for Charity, where Charlie Demerjian and I could turn a pile of parts into a working system under the pressure of competition. Jensen, consider the gauntlet thrown: when NVIDIA’s robots are ready for a proper charity PC build race, we’ll be waiting.
