PedSLight: pedestrian and safety-aware reinforcement learning for mixed pedestrian-vehicle signal control

Abstract

Signalized intersections with mixed vehicular and pedestrian traffic face persistent trade-offs between efficiency and safety, particularly under permissive left-turn control. This study develops adaptive signal control that embeds pedestrian conflict risk directly in the optimization objective rather than through heuristic constraints or phase restrictions. We propose PedSLight, a cycle-level reinforcement learning framework implemented in the Simulation of Urban MObility (SUMO) at a four-leg intersection and extended to a three-intersection corridor. Controllers observe direction-aware traffic states and select continuous cycle-level allocation ratios for protected and permissive left-turn phases. A threshold-based time-to-collision (TTC) surrogate cost is combined with waiting-time efficiency in a weighted reward trained with Proximal Policy Optimization (PPO). Performance is compared with actuated and fixed-time baselines under multiple demand levels using travel time and TTC-based safety metrics. At the single intersection, PedSLight achieves competitive travel times and a more favorable efficiency–safety trade-off than fixed-time control, with clearer advantages under higher demand. In the corridor, the learned multi-agent policy attains the lowest estimated average travel time and TTC-based surrogate conflict cost among the evaluated controllers. The framework enables explicit efficiency–safety trade-off control through stable cycle-boundary updates and supports simulation-based evaluation before deployment at pedestrian-intensive intersections with permissive left turns.