Structured Instruction Parsing and Scene Alignment For UAV Vision-Language Navigation

Recent advances in aerial Vision-and-Language Navigation (VLN) have introduced a more meaningful and practical paradigm of VLN by considering significantly longer paths and more complex spatial reasoning compared to ground-based VLN. However, the larger scale and increased complexity of outdoor environments in aerial VLN present substantial challenges in establishing accurate correspondence between textual instructions and visual scenes. In this work, we propose to incorporate Large Language Models (LLMs) to extract key components from navigation instructions and construct the corresponding subtasks. This structured instruction parsing module ensures the appropriate granularity of navigation instructions, enabling more precise alignment between language and visual cues. To further enhance the integration of multi-modal information and cross-modal understanding, we introduce a scene-based subtask alignment policy that effectively associates each parsed subtask with corresponding visual observations along the navigation path. Combined, the proposed approach significantly outperforms current state-of-the-art methods on the AerialVLN dataset.