NVIDIA open-sources ENPIRE, letting coding agents self-improve real robot policies

GuanyaShi · x · 2026-09-10

NVIDIA Research released ENPIRE, a harness where AI coding agents autonomously improve real-world robot policies via closed-loop physical trial, error, and evolution: agents propose changes, evaluate on a physical station, and inspect measurements to guide the next iteration. CMU also tested GPT-6 Astra with ENPIRE, achieving zero-shot robot in-context learning.

Original post →

More from coding & agent

coding & agent channel →