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

NeurIPS 2026 Evaluations & Datasets

Jiajun Jiang^1,* Chunliang Hua^1,* Zichun Chen² Yanxing Wu² Zeyuan Yang² Jie Song^1,3 Xiao Hu^1,2,†

¹ HKUST (Guangzhou)

² LASER, IDEA

³ HKUST

* Equal contribution † Corresponding author

Abstract

SatNav is a scalable benchmark for long-horizon UAV vision-language navigation built from high-resolution satellite imagery. It provides 118K episodes across 59 scenes in 18 cities, with an average trajectory length of 379 m. The Boundary, Landmark, and Route tasks test long-term memory and geospatial reasoning through language-guided navigation using local satellite crops as approximations of UAV nadir views. Evaluations reveal persistent challenges for classical VLN and large vision-language model agents. We also introduce SwiftVLN, a modular framework for comparing memory designs, and demonstrate that satellite-trained models can transfer to real-flight UAV observations.

Satellite imagery of the Maracanã stadium and its surroundings — Satellite scene
Rio de Janeiro

How are these scores calculated?

Stop within 30 m of the goal. For this loop, first move more than 60 m from the start, then return.

SR records whether you stop inside the goal radius. OS records whether you reach that radius at any step. Both show 0% or 100% for this single episode; the benchmark reports averages over many episodes.

SPL = SR × reference path length / max(reference path length, traveled length). A successful path that is no longer than the reference scores 1; longer paths score less. NE is your final distance to the goal in meters.

Dataset Statistics

Task Distribution

Trajectory Length

Dataset Splits

Benchmark Results

How well do navigation models generalize to new cities? We train or fine-tune all models on SatNav and evaluate them on seen and unseen scenes.

Model

SeenUnseen

* Initialized from a released navigation-trained checkpoint, then fine-tuned on SatNav. Unmarked LVLMs start from their base backbones.

View full results

SR, SPL and OS are percentages; NE is measured in meters. Steps is the average number of executed actions. Best SR, SPL, OS and NE values within each split are bold and underlined.
Model	Test Seen					Test Unseen
Model	SR ↑	SPL ↑	OS ↑	NE ↓	Steps	SR ↑	SPL ↑	OS ↑	NE ↓	Steps

SwiftVLN Framework

SwiftVLN combines a sliding dialogue window with long-term visual memory. Here, we show a window size of 4 and an overlap of 3.

One framework, interchangeable memories.

Choose which observations to retain.

Uniform: Samples evenly across the full history.
Random: Selects frames at random, then orders them by time.
Temporal-biased: Samples recent observations more densely.

Sampling details

Memory Module Ablation on SatNav

How does each memory design change navigation performance?

View full results

SR, SPL and OS are percentages. ΔSR and ΔOS are percentage-point changes relative to SwiftVLN reference. The largest gain within each ablation group and split is bold and underlined.
Memory design	Test Seen					Test Unseen
Memory design	SR ↑	SPL ↑	OS ↑	ΔSR ↑	ΔOS ↑	SR ↑	SPL ↑	OS ↑	ΔSR ↑	ΔOS ↑

Real-world UAV Flights

Trained on satellite imagery, SwiftVLN transfers to real-world UAV navigation with a lightweight visual adapter.

Boundary

Landmark

Route

See the paper appendix for deployment details and the Wiki for adapter training and integration.