SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery

2026-09-25

AI Summary

This study addresses the reliance of existing UAV vision-language navigation benchmarks on expensive 3D assets and their limited scalability to city-scale, long-horizon scenarios. To overcome these limitations, this work proposes a scalable benchmark and an automated data synthesis pipeline grounded in satellite imagery, yielding a large-scale dataset comprising 118,000 long-trajectory tasks. Methodologically, we introduce SwiftVLN, a modular framework that integrates large vision-language models with a modular memory mechanism to effectively tackle challenges in long-term memory retention and geographic reasoning. Experimental results demonstrate that the proposed approach reveals core difficulties inherent in urban-scale navigation while validating the zero-shot transferability and generalization capability of satellite-trained models to real-world UAV flight scenarios.

Abstract

Urban uncrewed aerial vehicle (UAV) vision-language navigation (VLN) requires agents to follow instructions across extended urban spaces, inherently demanding long-term memory and geospatial grounding. However, scaling existing benchmarks remains difficult because of their reliance on costly reconstructed 3D assets, limiting geographic diversity and episode scale. To address this, we introduce SatNav, a scalable, long-horizon UAV VLN benchmark built from high-resolution satellite imagery. SatNav targets city-level navigation missions and uses satellite crops as approximations of UAV nadir views for visual observations. Through an automated cue-to-episode pipeline, SatNav constructs 118K episodes from 59 scenes across 18 cities, with an average trajectory length of 379 m. To stress-test long-horizon memory and geospatial reasoning, SatNav defines three task families: Boundary, Landmark, and Route, targeting loop progress tracking, landmark-based spatial grounding, and route following with counting cues. Benchmarking classical VLN agents and recent agents based on large vision-language models (LVLMs) on SatNav shows that city-scale navigation remains challenging. We further introduce SwiftVLN, a modular framework with switchable memory components, and conduct systematic memory-design ablations. Finally, satellite-to-UAV transfer experiments show that satellite-trained navigation models can operate on real-flight UAV observations, showing the practical relevance of SatNav. Our project page: https://eku127.github.io/SatNav/

Problem

• Vision-Language Navigation
• UAV
• Satellite Imagery
• Long-Horizon
• Geospatial Grounding

Innovation

• Vision-Language Navigation
• Satellite Imagery
• Long-Horizon UAV
• Scalable Benchmark
• Memory Module