Dual-Branch Dynamic Perception and Interaction Framework for Aerial Vision-and-Language Navigation

Unmanned Aerial Vehicle (UAV) Vision-and-Language Navigation aims to enable drones to autonomously navigate complex environments using natural language instructions. Existing approaches often depend on static visual cues or isolated directional signals, thereby neglecting the dynamic interactions between visual perception and directional guidance, which limits their ability to capture cross-modal correlations. To address this challenge, we propose a Dual-Branch Dynamic Perception and Interaction framework, dubbed DBDP. Specifically, a time-aware visual branch is designed to dynamically capture sequential scene features, while a spatial search directional branch robustly extracts and interacts with navigational cues. These modality-specific features are then fused via a joint embedding module with language representations, yielding enhanced action prediction accuracy. Experimental results and ablation studies demonstrate that the proposed DBDP framework significantly outperforms state-of-the-art methods in terms of navigation performance and interpretability.