Nvidia is taking the next step in automating robot development. Researchers at the company have developed a system that allows AI agents to independently train, evaluate, and improve robots.
According to Ars Technica, in experiments, robots learned to perform tasks such as installing GPUs in motherboards, cutting tie wraps, and sorting small parts—all without requiring constant intervention from researchers.
The research comes from Nvidia’s GEAR lab (Generalist Embodied Agent Research) and was conducted in collaboration with Carnegie Mellon University and the University of California, Berkeley. At its core is a framework called ENPIRE that enables multiple AI coding agents to collaborate to improve robot performance.
Whereas robot training traditionally requires a great deal of manual work, ENPIRE allows the agents to design experiments, adapt training software, analyze results, and implement improvements on their own. The system combines various functions, including error analysis, automated validation of results, and parallel execution of experiments across multiple robots.
Robots Learn from Their Own Experiments
For the tests, the researchers used AI coding agents from various providers, including OpenAI, Anthropic, and Moonshot AI. The agents independently developed alternative training methods and compared the results in practice. Improvements that led to higher success rates were then retained for subsequent training rounds.
This yielded remarkable results. Robots succeeded in performing a variety of tasks at approximately 99 percent. In addition to installing GPUs into motherboards, these tasks included sorting pins into a storage box, securing and cutting through tie wraps, and performing standard manipulation exercises commonly used in robotics research.
According to the researchers, the AI agents even outperformed existing training methods, which involve human experts actively optimizing the learning process, on some tasks.
The experiments show that larger teams of AI agents make faster progress than individual agents. In a test setup, a group of eight agents achieved a 99% success rate within 2 hours. Four agents took about three hours to do the same, while a single agent needed nearly five hours to reach the same level.
However, these economies of scale are not unlimited. As more agents collaborate, the need for coordination among them also increases. The researchers observed that agents spent increasing amounts of time processing and summarizing each other’s findings. In addition, robots regularly came to a standstill while the AI systems analyzed code, investigated errors, or waited for responses from underlying language models.
A Step Toward Autonomous Robot Labs
The project aligns with Nvidia’s broader ambition to have AI not only develop software but also independently improve physical systems. According to Jim Fan, director of AI at Nvidia, part of the research work can now be fully automated. Researchers primarily review the results that the AI agents have produced overnight.
The researchers have announced that ENPIRE will be made available as open source. With this, Nvidia aims to enable universities, companies, and hobbyists to set up similar self-driving robot labs.
The research highlights how quickly the boundaries between generative AI and robotics are blurring. While AI coding agents were originally developed to support programmers, they are now increasingly being deployed as autonomous researchers who not only write software but also teach physical machines how to perform new tasks.