NVIDIA's GEAR lab unveiled ENPIRE, a software harness that lets AI coding agents control robotic arms to cut zip‑ties and insert GPUs into motherboard sockets. The framework wraps large language models with tool‑use, memory and feedback loops, enabling fully autonomous robot training. Researchers from Carnegie Mellon and UC Berkeley collaborated on the system and supports rapid debugging.

The team tested three agents—OpenAI’s Codex paired with GPT-5.5, Anthropic’s Claude Code on Opus 4.7, and Moonshot AI’s Kimi Code using Kimi K2.6. Each generated its own training algorithms, ran parallel experiments on multiple robots, and retained refinements that boosted success rates across cycles. The framework also records sensor data for post‑run analysis. Results appeared in a paper released June 16, 2026.

Jim Fan, NVIDIA’s AI director, posted that the ENPIRE lab now “self‑improves tirelessly overnight,” suggesting a shift toward hands‑free hardware provisioning. By open‑sourcing the harness, the researchers aim to let hobbyists spin up personal robot labs without manual coding. Such capability could streamline server‑rack assembly, where precise GPU placement is time‑critical.