AirUniNav: Unified Vision-Language Navigation for UAVs in Indoor and Outdoor Scenes

Vision-language navigation(VLN) for unmanned aerial vehicle(UAV) represents a pivotal challenge in embodied intelligence, bridging perceptual understanding with complex decision-making. Despite rapid progress, existing research often treats indoor and outdoor environments in isolation, ignoring the intrinsic visual and scale discrepancies between these environments. Consequently, current models lack the adaptability needed for real-world deployment. To bridge this gap, we present AirUniNav, a unified framework designed to master both precise indoor operations and large-scale outdoor navigation. AirUniNav enables cross-domain collaborative learning through the construction of a large-scale joint indoor-outdoor dataset and the design of a scenario-adaptive training strategy. To reduce redundant visual processing and improve efficiency, we design a dynamic frame sampling mechanism with time-decay weighting to reduce computational cost, and a vision-text token interaction module to enhance temporal modeling. Extensive experiments demonstrate that AirUniNav achieves superior generalization and navigation efficiency across diverse indoor and outdoor tasks, highlighting its potential to advance UAV VLN toward real-world applications.