NVIDIA researchers have developed ENPIRE, a new software framework enabling AI coding agents to autonomously direct robot training. Developed by the NVIDIA GEAR lab alongside Carnegie Mellon University and UC Berkeley, the system allows agents to manage compute resources and execute complex physical tasks like inserting GPUs into motherboards.
Jim Fan, Director of AI at NVIDIA, confirmed the lab now self-improves overnight while engineers review results in the morning. The team plans to open-source the entire framework, allowing external developers to host self-running robot labs. This move aims to democratize advanced robotics research beyond major tech hubs.
ENPIRE operates through four core modules that handle automatic resets, policy refinement, parallel evaluation, and failure analysis. The system ingests research papers and analyzes logs to continuously improve training infrastructure and algorithm code without human oversight.
Testing validated the framework using three distinct AI agents: OpenAI Codex with GPT-5.5, Anthropic Claude Code with Opus 4.7, and Moonshot AI Kimi Code. Each agent independently developed algorithmic approaches, retaining only changes that increased success rates during repeated cycles of self-directed physical experimentation.