Nvidia's Robot Fleet Out-Trains Itself While Humans Just Write the Paper 📎

Nvidia, Carnegie Mellon University, and UC Berkeley researchers published a paper Tuesday detailing ENPIRE, a framework that lets AI coding agents train physical robots without human supervision. The system was tested on eight bimanual robot arms at Nvidia's GEAR lab, which spent the past few weeks teaching themselves to insert pins, seat graphics cards, and cut zip ties.

The setup splits work into two stages. A human first helps the agent build a reset routine that returns the workspace to a clean starting position and a reward function that scores success from camera footage. After that, the agent takes over completely, searching published research for ideas, choosing between training methods such as imitation learning, reinforcement learning, or hand-written rules, and rewriting its own code before testing on the robot. Coding agents cited in the paper include OpenAI's Codex, Anthropic's Claude Code, and Moonshot's Kimi Code.

The eight stations ran independently on their own hardware and computers, syncing progress through Git so a winning idea could spread across the fleet within minutes. Scaling from one robot to eight cut the time to master the Push-T benchmark, in which a robot slides a T-shaped block into a target zone using only pushes, from roughly five hours to two. Pin insertion into 4-millimeter holes dropped from more than 90 minutes to about 40. Across four real-world tasks, the agents drove policies to a 99% success rate, according to the paper, reaching near-perfect reliability on pin insertion faster than a comparable human-in-the-loop method.

Jim Fan, GEAR Lab co-lead and director of Nvidia's AI research, called the project an effort to enable AutoResearch in the physical world for the first time, saying the team handed the agents a fleet of robots, a GPU allocation, and a token budget, then stepped back.