Aeroseek: Exploring LLM-Enhanced Monte-Carlo Path Planning for UAV Navigation

While Unmanned Aerial Vehicles (UAVs) hold immense potential for transforming the low-altitude economy, developing an agent that combines both intelligence and flexibility remains a significant challenge. To this end, we propose a novel aerial navigation system, Aeroseek, to search for executable paths that coherently transition from cognitive understanding to physical movement, enabling more autonomous and flexible navigation with instruction guidance. Specifically, given a navigation instruction, we follow a fast-andslow thinking procedure inspired by human pilots by breaking down complex decisions into manageable sub-tasks and initiating motion planning immediately upon the completion of each sub-task for realtime movement. By integrating a multi-branch Chain-of-Thought, Aeroseek is capable of combining reasoning with spatial recognition and self-verification to search for robust sub-goals under expected constraints. For sub-goal navigation, we introduce Online Kinodynamic Monte-Carlo Path Planning (OnKi-MCPP), a novel motion planning approach handling both kinodynamic constraints and path search, while ensuring obstacle avoidance for UAVs. Benefiting from the proposed two-tier searching protocol, Aeroseek demonstrates comprehensive capabilities in processing complicated instructions and adapts to dynamic aerial conditions. Extensive experiments demonstrate the advantages of Aeroseek across tasks of varying complexity.