Abstract
Efficient fire drill evacuation in high-rise buildings demands real-time perception and optimized path planning. This study presents an autonomous evacuation framework that integrates deep learning-based object detection with multi-objective trajectory planning. Using RGB-D vision, the system detects human traffic and staircases, projecting them into the robot's workspace. The evacuation task is modeled as a Moving Target Traveling Salesman Problem (MT-TSP) and solved via a modified NSGA-II algorithm, balancing travel time, path length, safety, and accessibility. Experiments show over 90% detection accuracy and significant improvements in navigation efficiency, demonstrating the effectiveness of combining deep learning with evolutionary optimization for robotic fire drill planning. Note to Practitioners-This work addresses the challenge of autonomous robot-assisted evacuation in dynamic fire scenarios, where multiple moving individuals must be safely guided to exits. the work presents a multi-objective evolutionary path planning framework that integrates real-time perception to detect and track moving targets. By optimizing for travel time, safety, and target reachability using NSGA-II, the system enables robots to make informed navigation decisions under uncertainty. Practitioners developing autonomous emergency response systems can adapt this approach for real-world applications involving human tracking, multi-criteria navigation, and dynamic environment awareness.