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

The article introduces ENPIRE, a framework designed to enhance robotic manipulation through a structured feedback loop enabling autonomous policy improvement in real-world settings. ENPIRE consists of four modules: Environment (for scene reset and verification), Policy Improvement (for refining policies), Rollout (for evaluating policies with multiple robots), and Evolution (for analyzing performance and enhancing algorithms). This system enables coding agents to achieve a 99% success rate on complex tasks, significantly reducing human intervention and offering a scalable method for advancing robotics research and applications.