| description abstract | Drones are becoming indispensable in emergency search and rescue (SAR), particularly in intricate urban areas where rapid and accurate response is crucial. This study addresses the pressing need for enhancing drone navigation in such complex, dynamic urban environments, where obstacles like building layouts and varying wind conditions create unique challenges. Particularly, the need for adapting drone autonomous navigation in correspondence with dynamic wind conditions in urban settings is emphasized because it is important for drones to avoid loss of control or crashes during SAR. This paper introduces a pioneering method integrating multiobjective reinforcement learning (MORL) with a convolutional autoencoder to train autonomous drones in comprehending and reacting to aerodynamic features in urban SAR. MORL enables the drone to optimize multiple goals, whereas the convolutional autoencoder generates synthetic wind simulations with a substantially lower computation cost compared to traditional computational fluid dynamics (CFD) simulations. A unique data transfer structure is also proposed, which fosters a seamless integration of perception and decision-making between machine learning (ML) and reinforcement learning (RL) components. This approach uses imagery data, specific to building layouts, allowing the drone to autonomously formulate policies, prioritize navigation decisions, optimize paths, and mitigate the impact of wind, all while negating the necessity for conventional aerodynamic force sensors. The method was validated with a model of New York City, offering substantial implications for enhancing automation algorithms in urban SAR. This innovation enables the possibility of more efficient, precise, and timely drone SAR operations within intricate urban landscapes. | |
