NVIDIA’s New AI System Lets Robots Teach Themselves Overnight

A team of researchers has built a software framework that lets AI coding agents take over the job of training physical robots—with no human supervision. The system, called ENPIRE, was developed by scientists at NVIDIA’s GEAR lab, Carnegie Mellon University, and UC Berkeley. When given access to robotic arms, computing power, and a generous token budget, the AI agents figured out how to teach robots to cut zip ties and insert GPUs into thin motherboard sockets.

“A part of our NVIDIA GEAR lab now self-improves tirelessly overnight,” wrote Jim Fan, NVIDIA’s AI director, on LinkedIn. “We just read the reports in the morning.” He joked that the goal was for everyone to take a holiday without CEO Jensen Huang noticing. The team plans to open-source the framework, so anyone can run their own self-training robot lab at home.

ENPIRE is built around four modules. It automatically resets and verifies tasks, refines the policies that guide robot behavior, evaluates those policies across multiple robots working in parallel, and fixes failures by analyzing logs, reading research papers, and improving training code. The researchers tested ENPIRE with three different AI coding agents: OpenAI’s Codex with GPT-5.5, Anthropic’s Claude Code with Opus 4.7, and Moonshot AI’s Kimi Code with Kimi K2.6. Each team of agents independently developed its own training approach, ran real-world experiments, and kept whatever changes boosted the overall success rate across repeated cycles of self-directed testing.

The full technical paper was published on June 16, 2026.