AeroGuide: Grounded Vision-Language Navigation in Aerial Environments

Aerial vision-language navigation (VLN) requires a UAV to follow natural language instructions from visual observations. Aerial VLN is more challenging than ground-level VLN because UAV observations cover large outdoor areas, landmarks often occupy only a small part of the image, and navigation requires complex spatial reasoning. General-purpose vision-language models (VLMs) often struggle to identify instruction-relevant visual evidence from such dense aerial observations. We propose AeroGuide, a grounded aerial VLN framework built on Qwen3-VL. AeroGuide introduces Navigation-conditioned Visual Token Routing (NVTR), which selects instruction-relevant visual tokens conditioned on the current observation, UAV pose, and temporal landmark memory. AeroGuide maintains an egocentric temporal landmark memory that retains compact landmark evidence rather than full historical frames. The model is trained through multiple phases, including supervised learning, reasoning distillation, and simulator-based policy optimization, to improve grounded action prediction and closed-loop navigation. On the OpenFly benchmark, AeroGuide improves navigation success rate from 34.3% to 45.1%, a 31.5% relative improvement over the best-performing prior baseline. Ablation studies show that explicit visual grounding and compact temporal memory are key contributors to this improvement, demonstrating the effectiveness of grounded navigation for aerial VLN. Project page is available at https://aeroguide-frontend-iota.vercel.app/.