Nvidia researchers use AI coding agents to run robot training tests

In short: Nvidia-backed researchers built ENPIRE, software that lets AI coding agents run many robot training experiments on their own.

What happened

Researchers at Nvidia’s GEAR lab, with collaborators at Carnegie Mellon University and UC Berkeley, introduced an “agent harness” called ENPIRE. Think of it like a manager layer that helps an AI model use tools, remember past attempts, follow rules, and try again with feedback (like running a tight practice routine).

Using ENPIRE, teams of AI coding agents wrote and updated training code for real robotic arms. The robots learned tasks such as moving a T-shaped block into place, organizing pins, tying and cutting zip ties, and inserting and removing a GPU from a motherboard slot.

In tests, the researchers reported up to a 99 percent success rate across several tasks. They also found that larger teams, up to eight agents, often reached high success rates faster than smaller teams. For example, an eight-agent team hit 99 percent success on the “Push-T” task in about two hours of research time, versus nearly five hours for a single agent.

Why it matters

Training robots usually takes many rounds of trial and error, and people often have to watch, reset, and tweak the software. Systems like ENPIRE aim to automate more of that work, which could speed up how quickly labs teach robots new physical skills.

The study also highlights practical limits. Robots sometimes sat idle while the agents read logs, debugged code, or waited on the underlying AI model. More agents also used more “tokens” (the paid units of text an AI service processes), which can raise costs.

Source: Arstechnica