LensDesigner is an autonomous agent that automates optical lens design, a complex task traditionally requiring expert human intuition. It addresses the cold start problem by constructing LensLib100K, a large library of lens configurations, and using Optics-Aware Retrieval to provide physically valid initial seeds. The agent then performs macro orchestration within a simulated environment, receiving real-time optical feedback.
To continuously improve, LensDesigner incorporates a self-evolving mechanism guided by a curriculum agent that gradually increases task difficulty. This allows the agent to extract, accumulate, and reuse design heuristics over time. Evaluated on the new LensArena of 120 diverse design tasks, the system significantly outperforms existing baseline algorithms in success rate and optimization efficiency, demonstrating a step toward intelligent, automated lens design.