A Queue-Aware Multi-Agent Eco-Driving Model for Connected and Automated Vehicles at a Signalized Intersection in Mixed Traffic

Eco-driving at signalized intersections uses signal timing to plan vehicle arrivals, smooth speed changes, and reduce CO₂ emissions. However, the effectiveness of eco-driving policies may be limited without considering queue discharge, because queues ahead can delay vehicle passage even during a green phase. This study proposes a queue-aware multi-agent eco-driving model for connected and automated vehicles (CAVs) in mixed traffic. First, an effective green window is estimated for each CAV by accounting for queue discharge. Second, a local-emission reward combines emissions from the controlled CAV and nearby same-lane followers. Finally, these components are integrated into a proximal policy optimization controller that generates continuous acceleration commands. The proposed model is evaluated in SUMO across varying traffic demands and CAV penetration rates. Results show that CO₂ emissions reduce in most tested conditions and the reduction rates grow with CAV penetration at each demand level. At 1200 veh/h, the reduction rises from 7.0% to 25.6% as CAV penetration increases from 10% to 70%. Ablation results further reveal that the effective green window and local-emission reward jointly contribute to the CO₂ reductions. These findings highlight the potential of the proposed method for eco-driving in mixed traffic.