Multi-UAVs cooperative task allocation and path planning for low-altitude logistics

To address the challenges of collaborative task allocation and path planning for multiple logistics unmanned aerial vehicles (UAVs) in urban low-altitude environments, this paper proposes a bilevel nested joint optimization method based on reinforcement learning and a graph search algorithm to enhance the efficiency of collaborative last-mile delivery by multiple logistics UAVs while reducing flight risks. The proposed method constructs a bilevel architecture system based on a task allocation and decision-making model and a path planning model. The upper-level model holistically considers the demands of three stakeholders—government (safety), customers (timeliness), and UAV enterprises (economy)—at the macro level. Then, based on real-time order information and UAV status, a multi-objective optimization and constraint model is constructed under complex dynamic environments. A multi-agent proximal policy optimization algorithm is employed to achieve rapid dynamic task allocation and decision-making. The lower-layer model utilizes the upper-level allocation results combined with detailed environmental information to plan safe and efficient flight paths for each UAV at the micro level. It employs an improved jumping-point search algorithm for refined path optimization. A loop feedback mechanism is designed to facilitate information exchange between layers, thereby coupling the task allocation and path planning processes to achieve collaborative optimization of upper- and lower-level task allocation and decision-making. This method effectively addresses complex logistics delivery scenarios, enhancing the overall efficiency and robustness of the delivery system. Simulation experiments comprehensively consider path influences from flexible open-area delivery, varying numbers of distribution centers and UAVs, and on-demand rush orders. Tests conducted in medium- and high-density environments demonstrate the proposed model and algorithm’s significant superiority in dynamic complex scenarios. Even when confronted with complex environments and dynamic order scenarios, it consistently generates highly applicable UAV flight paths.