Diffusion-Head VLA Model for Adaptive UAV Vision-Language Navigation

Visual-Language-Action (VLA) models demonstrate strong semantic understanding and action modeling capabilities in UAV Visual-Language Navigation (VLN) tasks, yet they perform ineffectively in complex scenarios, such as long-distance and multi-dimensional continuous motion tasks. To address this, we propose a Diffusion-Head VLA model with an Adaptive Action Policy (AAP), enabling continuous and fine-grained action generation. This strategy allows the model to leverage high-level semantic guidance from Vision-Language Models (VLMs) while producing continuous and fine-grained control via the diffusion head. The Adaptive Action Policy dynamically integrates the actions generated by the VLM and the diffusion-based action head, thereby improving the UAV's performance. The History-Aware Directional Perturbation (HADP) mechanism injects perturbations aligned with historical action trajectories, enhancing both action continuity and diversity, which leads to more stable and effective training. Experimental results compared with baselines show that our approach improves UAV navigation performance.