Nvidia's ENPIRE lets AI agents train robots to install GPUs and cut zip ties

The ENPIRE framework, created by Nvidia's GEAR lab in collaboration with Carnegie Mellon and UC Berkeley, wraps around AI models to give them memory, context, and feedback loops. It equips coding agents with four modules: automatic reset/verification, policy refinement, parallel evaluation across multiple robots, and failure analysis that includes ingesting research papers. In tests, agents using OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot's Kimi Code with K2.6 independently developed training strategies for tasks like Push-T (99% success), pin organization, zip tie cutting, and GPU insertion. Larger agent teams (up to eight) achieved success faster, but at higher token costs. The system is open-sourced, enabling anyone to host a self-running robot lab.

Despite promising results, limitations emerged: robots sat idle while agents debugged or summarized ideas, and larger teams consumed more tokens without increasing robot utilization. However, on the pin organization task, AI agents achieved near-100% success faster than a human-in-the-loop method. The ENPIRE harness allows the Nvidia GEAR lab to self-improve overnight, with director Jim Fan noting, 'We just read the reports in the morning.' The work demonstrates a step toward fully autonomous robot training, though token costs and idle time remain challenges for widespread adoption.