Ep#110: ENPIRE: Agentic Robot PolicySelf-Improvement in the Real World

The ENPIRE project introduces a framework for agentic robot policy self-improvement in real-world environments, addressing the bottleneck of human intervention in robotics research. By decomposing the system into modules for resetting, reward function generation, and policy optimization, the approach allows coding agents to autonomously iterate on physical tasks like pin insertion and object manipulation. These agents leverage visual-language models to interpret camera feedback, synthesize reward functions, and execute motion planning, effectively reducing the need for manual teleoperation and heuristic tuning. While current implementations operate under limited compute budgets, scaling the number of agent teams and utilizing advanced models like Astra enhances performance and enables complex behaviors such as in-context learning and object handovers. This paradigm shifts the role of roboticists from low-level implementation to high-level system design and verification, as autonomous agents increasingly handle the iterative trial-and-error processes required for real-world robotic development.