ENPIRE: Agentic Robot Policy Self-Improvement in the Real World

ENPIRE wraps a coding agent around a four-module physical feedback loop — automatic scene reset, policy execution, outcome verification, and iterative refinement — enabling autonomous robot policy improvement without human supervision between trials. The framework targets the gap between LLM-based algorithm search, which works in simulation, and real-world dexterous manipulation where environment resets have historically required human intervention.