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

SatNav. A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery. By Jiajun Jiang and six others, from the Hong Kong University of Science and Technology, Guangzhou, the International Digital Economy Academy, and one other institution. Published on the twenty-fifth of September. Its code repository is worked out: about two hundred source files, with a readme and a licence.

A drone navigating a city by following spoken or written instructions has to cover long distances, and that demands two capacities: a long-term memory of where it has been, and an ability to ground its position geographically. Existing benchmarks for this kind of vision-language navigation are hard to scale. They depend on reconstructed 3D assets that are costly to make, and that cost limits both how many places they cover and how many episodes they can offer.

SatNav takes a different source. It is a scalable, long-horizon benchmark for navigating drones, built from high-resolution satellite imagery. Its missions are at the level of a whole city, and crops of satellite images stand in as approximations of what a drone sees looking straight down. To stress memory and geospatial reasoning over long horizons, it defines three task families. Boundary targets tracking progress around a loop. Landmark targets spatial grounding by reference to landmarks. Route targets following a route with cues that involve counting.

Episodes come from an automated pipeline that turns cues into episodes. With it, SatNav builds one hundred and eighteen thousand episodes from fifty-nine scenes, spread over eighteen cities. The average trajectory runs three hundred and seventy-nine metres. Because the imagery is satellite data rather than handmade reconstruction, adding geography is a matter of adding imagery.

The authors benchmark classical navigation agents alongside recent ones built on large vision-language models, and the result is that navigation at city scale remains challenging. They also introduce SwiftVLN, a modular framework whose memory components can be switched in and out, and use it to run systematic ablations of memory design. A final set of experiments tests transfer from satellite to drone. Navigation models trained on satellite views can operate on observations from real flights, which the authors take as evidence of the benchmark's practical relevance.

The bet underneath SatNav is that the view from orbit is close enough to the view from a drone to train on. If that holds, the bottleneck of expensive 3D reconstruction gives way to imagery that already exists, and long-horizon navigation can be tested at a scale that matches the cities it is meant for.