Beyond Traffic Signals: Coordinating Heterogeneous Mixed Traffic at Unsignalized Intersections

Urban intersections are inherently complex, hosting a diverse mix of vehicles ranging from small cars to large semi-trailers, each with distinct driving behaviors and space requirements. Managing such heterogeneous traffic becomes particularly challenging at unsignalized intersections, where the absence of traffic lights demands real-time coordination. This thesis explores how robot vehicles (RVs), powered by reinforcement learning (RL), can dynamically optimize mixed traffic flow under varying degrees of automation. By gradually increasing RV penetration from 10% to 90%, our findings reveal a substantial reduction in average waiting times—up to 86% compared to signalized intersections—demonstrating the efficiency of RL-based control. Notably, we observe a "rarity advantage", where less frequent vehicle types, such as trucks, experience the most significant improvements, benefiting from RV-driven coordination by as much as 87%. Additionally, space headways decrease consistently across all vehicle types, suggesting enhanced road space utilization. While RVs operate at higher speeds, leading to increased energy consumption, the overall efficiency gains surpass those of conventional traffic signals. These insights underscore the transformative potential of RL in coordinating heterogeneous mixed traffic, paving the way for more adaptive and scalable traffic management solutions in real-world urban mobility.