Robot-assisted pedestrian evacuation in fire scenarios based on deep reinforcement learning

Abstract

Indoor fires pose a significant challenge to the safe evacuation of pedestrians. In response to fire hazards, pedestrians instinctively seek alternative evacuation routes to avoid the hazard zone. However, the specific location and intensity of the fire hazard zone can influence pedestrians' decisions, leading to varying congestion levels in different areas. To address this challenge and enhance overall evacuation efficiency, this paper introduces an improved social force model to depict pedestrian movement in fire scenarios and proposes a methodology that leverages dynamic robot for pedestrian evacuation, employing Deep Reinforcement Learning (DRL) and Human-Robot Interaction (HRI). The results show that in the no-robot scenario, pedestrians will detour according to the varying locations of fire hazard zones and emergency levels, resulting in congestion at different positions. In the static robot scenario, robots placed in different locations exhibit varied effects on evacuation depending on the fire hazard zones' locations and intensities. In the DRL-control robot scenario, the robot controlled by DRL and HRL can always navigate to the appropriate position to promote evacuation, regardless of the fire's location and emergency levels or the robot's initial placement. Furthermore, our findings reveal that strategically positioned robots can enhance evacuation efficiency by alleviating crowding and increasing the distance between pedestrians and fire hazard zones in most cases, thereby improving pedestrian safety. This study offers practical guidance for managing pedestrian evacuation during fire incidents and establishes a theoretical foundation for refining evacuation strategies and safety measures at fire scenes.