PCBWorld: New Benchmark for AI-Driven PCB Design Automation

Automating the routing of Printed Circuit Boards (PCBs) with AI has been challenging, with learning-based methods often falling short of traditional rule-based routers. To bridge this gap, a new open-source environment called PCBWorld has been introduced. This environment is built on the KiCad EDA engine, allowing AI agents to interact with the PCB design process in a way similar to human engineers. Agents within PCBWorld can perform native KiCad operations and receive real-time feedback from the Design Rule Check (DRC) system, ensuring that their routing adheres to strict design rules. The environment supports both reinforcement learning (RL) policies and tool-using large language model (LLM) agents. Alongside the environment, PCBWorld-Bench provides a comprehensive dataset of both controllable synthetic instances and 679 real open-source boards in KiCad's native format. The benchmark evaluates completed boards using eight engine-checked metrics. Initial experiments showed that agents trained in PCBWorld consistently outperformed simpler RL and open-loop LLM baselines. Notably, an RL policy trained solely on synthetic boards successfully transferred its knowledge to real boards without further training, achieving performance close to rule-based routers. This positions PCBWorld as a significant step forward for AI in PCB design automation.