DeCoR: Reinforcement Learning Co-Optimizes Urban Street Design and Traffic Control

A new two-stage reinforcement learning framework named DeCoR has been developed by researchers to optimize both crosswalk designs and network-level signal management for urban roadways based on flow observations. The initial design phase represents the pedestrian network as a graph, creating a generative policy that defines a Gaussian mixture model for crosswalk locations and widths, enabling the sampling of new crosswalks. For each proposed layout, a unified control policy adapts signal timings to reduce delays for both pedestrians and vehicles. In a 750-meter urban corridor, DeCoR achieved a 23% decrease in pedestrian arrival time at crosswalks while utilizing fewer crosswalks than current setups. Additionally, it cut pedestrian wait times by 79%. This research is detailed in the arXiv paper titled "DeCoR: Design and Control Co-Optimization for Urban Streets Using Reinforcement Learning" (arXiv:2605.21311).