Thunderbird: Adaptive Fast-Slow UAV Navigation with Scene Complexity Awareness

Recent UAV-VLN methods often employ complex VLM reasoning pipelines to improve navigation reliability, but apply them uniformly at every decision step—even in simple or unambiguous states where lightweight reasoning would suffice—resulting in unnecessary latency and computation. To address this, we propose a fast–slow navigation framework that separates high-level waypoint decision-making from low-level flight execution. At every navigation decision step, the slow system generates an intermediate waypoint and dynamically adjusts its VLM reasoning budget according to the difficulty of the current navigation state. The fast system then performs waypoint-to-waypoint motion planning, converting the selected waypoint into a safe and executable trajectory. To coordinate the fast and slow systems, we introduce the OVL scene complexity score that estimates the difficulty of the current navigation decision. Experimental results show that, compared with a conventional slow-system baseline that performs full reasoning at every navigation decision step, our method reduces the average VLM reasoning time by 49.4% and the flight time by 65.2%. Furthermore, compared with representative state-of-the-art methods, our approach achieves 2.57$\times$ the flight speed without compromising the navigation success rate.