ENPIRE Framework Enables Autonomous Robot Policy Self-Improvement

A new research paper introduces ENPIRE, a novel framework designed to facilitate autonomous policy self-improvement for robots operating in physical environments. Traditionally, achieving precise robotic manipulation in the real world has demanded extensive human oversight and intricate algorithm engineering, posing a significant hurdle to developing general physical intelligence. While coding agents have shown promise in automating algorithm search, their application has largely been confined to digital simulations. ENPIRE addresses this limitation by establishing a repeatable feedback loop for real-world policy refinement. This framework comprises four core modules: an Environment module for automatic scene reset and outcome verification, a Policy Improvement module for launching policy refinements, a Rollout module for parallel evaluation of policies on physical robots, and an Evolution module where coding agents analyze data, consult literature, and enhance training infrastructure and algorithms to resolve failures. This closed-loop system transforms robot learning into an optimized, human-effort-minimized process. Leveraging ENPIRE, advanced coding agents have demonstrated the ability to autonomously train policies, achieving a 99% success rate on challenging dexterous manipulation tasks, including organizing objects, fastening zip ties, and using tools. The efficiency of this process further increases when a team of agents is deployed across a fleet of robots. These findings suggest a scalable and practical pathway for deploying coding agents to independently advance robotics in the physical world.