Nvidia Builds Robots That Train Themselves Using AI Coding Agents

AI researchers at Nvidia have unveiled ENPIRE, a framework that allows coding agents to autonomously train and improve robots on real physical hardware. The system marks the first time artificial intelligence agents have conducted what researchers call "physical autoresearch," running the entire scientific loop from experiment design to hardware testing without human intervention. The research is a collaboration between Nvidia, Carnegie Mellon University, and UC Berkeley.

How ENPIRE Trains Robots Automatically

ENPIRE, which stands for Environment, Policy Improvement, Rollout, and Evolution, connects large language models (LLMs) — the same type of AI technology powering chatbots like ChatGPT — directly to physical robots. The framework breaks the robotics research process into four modules that work together in a continuous loop. The Environment module handles automatic scene resets and verification. Policy Improvement generates and revises robot behavior code based on past results. Rollout executes physical trials across entire robot fleets. Evolution compares different approaches and eliminates failing strategies while keeping what works.

What makes ENPIRE unique is that AI coding agents do all the heavy lifting. The team tested three leading tools: OpenAI’s Codex, Anthropic’s Claude Code, and Moonshot’s Kimi Code. These agents analyze performance logs, propose algorithmic improvements, test their hypotheses on real robots, and iterate on results — all without waiting for a human researcher to step in. The system can run experiments in parallel across multiple machines, speeding up learning.

Near-Perfect Results on Demanding Tasks

The results speak for themselves. Robot teams powered by ENPIRE achieved a 99 percent success rate on demanding manipulation tasks. These included installing a graphics processing unit (GPU) into a computer motherboard, cutting zip ties with precision tools, and inserting pins into designated slots. Each task requires the kind of careful, contact-rich movement that has traditionally been extremely difficult for robots to learn.

The framework scaled from single robots to eight-robot fleets working simultaneously. Nvidia plans to release ENPIRE as open source, which would allow universities, companies, and hobbyists to set up similar self-improving robot labs.

ENPIRE represents a significant shift in how robotics research could work. Instead of scientists spending months fine-tuning robot behaviors — the process of adjusting an AI model to perform specific tasks better — coding agents can now run that loop continuously. While challenges remain with computing efficiency during reasoning phases, the framework opens the door to robot labs that improve themselves around the clock.