Safety Prioritized Reinforcement Learning Enabled Collision Reduction in A 3D City Traffic Simulation Environment

Traffic collisions impose substantial costs in human lives, lost productivity, and economic burden while creating cascading civic, environmental and social challenges worldwide. Traditional real world signalized intersections utilizing fixed interval, actuated or adaptive traffic signals have shown limited success in addressing these complex, dynamic real-time scenarios. Current traffic simulation environments are unable to model realistic collisions directly and typically use surrogate measures of risk and conflict. While Reinforcement Learning (RL) models have demonstrated improved congestion over traditional traffic signal control, they have rarely addressed collisions. To address the current research gaps, three potential tools are developed: a comprehensive 3D city-wide simulation environment that integrates both macroscopic and microscopic traffic dynamics; a direct physics-informed collision model via Unity game engine; and a custom reward RL framework prioritizing safety and fatality reduction. There was a 31.79% reduction in serious collisions and 37.75% reduction in vehicle to non-vehicle collisions. Customized reward shaping allows balancing risk tolerance with efficiency in this larger scale traffic simulation incorporating both micro and macroscopic traffic flow features, in turn, demonstrating feasibility of applications incorporating the vision-zero safety principles of the Department of Transportation.