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

Jiajun Jiang; Chunliang Hua; Zichun Chen; Yanxing Wu; Zeyuan Yang; Jie Song; Xiao Hu

cs.CVSep 25, 2026

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

Authors: Jiajun Jiang, Chunliang Hua, Zichun Chen, Yanxing Wu, Zeyuan Yang, Jie Song, Xiao Hu

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Low Altitude Space Economy Research Center, International Digital Economy Academy (IDEA) · The Hong Kong University of Science and Technology

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/

Figures & tables

Figure 1: Overview of SatNav . The benchmark spans diverse cities and scenes, and introduces three task families: Boundary ( partial-arc , extended-loop , full-loop ), Landmark ( one-turn , two-turn ), and Route ( road , waterway , hybrid ). During navigation, the agent receives an instruction and a satellite crop generated by rotate-and-crop operations, then predicts and executes actions in the satellite map.

Figure 2: Overview of the SatNav episode-generation pipeline. Structured geospatial cues for the three task families are extracted from aligned satellite imagery and OSM annotations and then used to generate long-horizon trajectories and cue-grounded navigation instructions.

Panel of Figure 3 — Figure 3: SatNav episode-level statistics

Figure 4: Overview of the SwiftVLN framework.

Model	Test Seen					Test Unseen
Model	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps
Seq2Seq	2.1	2.1	31.1	180.36	53.29	1.6	1.5	29.6	219.17	58.50
CMA	9.5	9.3	44.9	323.25	94.85	7.3	7.2	43.9	363.59	99.31
OpenFly	13.2	13.0	33.5	166.88	58.93	11.7	11.6	32.2	195.91	66.35
OpenFly ∗	21.1	21.0	38.1	163.07	57.85	17.1	16.9	34.9	196.90	65.56
NaVILA	18.1	18.0	27.6	93.05	48.08	13.0	12.7	23.7	128.88	53.71

Table 1: Performance comparison on SatNav. SR, SPL, and OS are reported as percentages. Best results are in bold , and second-best results are underlined .

Memory Design	Test Seen					Test Unseen
Memory Design	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	$\Delta$ SR $\uparrow$	$\Delta$ OS $\uparrow$	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	$\Delta$ SR $\uparrow$	$\Delta$ OS $\uparrow$
Memory necessity
SwiftVLN reference	65.8	65.5	72.4	0.0	0.0	53.7	53.2	64.1	0.0	0.0
Short-term only	44.5	43.8	62.8	-21.3	-9.6	32.4	31.5	51.4	-21.3	-12.7
Short-term sliding window
Overlap turns( $N_{o}=2$ )	68.7	64.0	81.3	+2.9	+8.9	56.8	53.7	70.1	+3.1	+6.0

Table 2: SwiftVLN ablation study of memory designs on SatNav. SR, SPL, and OS are reported as percentages, while

\Delta

columns report absolute percentage-point changes relative to SwiftVLN.

Figure 5: A real-world Boundary task around a building with a grey roof.

Method	Boundary SR	Landmark SR	Route SR	Overall SR
SwiftVLN	33.3	33.3	50.0	38.9
SwiftVLN w/ adapter	33.3	50.0	83.3	55.6

Table 3: Real-world navigation performance of SwiftVLN with and without the visual adapter. Each setting is evaluated on 18 episodes, with six per task family. SR is reported as a percentage.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

City/Region	Scenes	Avg. Area	Episodes	City/Region	Scenes	Avg. Area	Episodes
	(#)	(km 2 )	(#)		(#)	(km 2 )	(#)
Amsterdam (EU)	2	6.56	6,424	Minneapolis (NA)	6	11.52	10,523
Auckland (OC)	1	15.17	2,391	New York (NA)	5	16.42	10,519
Berlin (EU)	5	11.58	8,942	Orlando (NA)	1	47.73	3,277
Boston (NA)	5	14.14	11,954	Paris (EU)	5	11.55	7,892
Brugge (EU)	1	8.75	2,845	Rio de Janeiro (SA)	1	22.24	1,975

Appendix

Table 4: Statistics of scenes and episodes in the SatNav dataset.

	Scale					Diversity
Dataset	Cov. ( $\mathrm{km}^{2}$ )	$N_{\text{scene}}$	$N_{\text{ep}}$	Traj. Sum ( $\mathrm{km}$ )	Avg. Len.	Traj. Types	Vocab.	Lex. Div.
AVDN	642	–	3.1K	879	289	2	2.2K	0.267
OpenFly	258	14	100K	9.91K	99	1	6.4K	0.247
AerialVLN	–	25	25.3K	16.77K	662	1	7.0K	0.246
HUGE-Bench	6.45	4	6.2K	2.56K	415	8	346	0.087
OpenUAV	–	22	12.1K	3.10K	255	1	–	–

Appendix

Table 5: Comparison of aerial navigation benchmarks in terms of scale and diversity. Coverage area is reported in

\mathrm{km}^{2}

, and total trajectory length is reported in

\mathrm{km}

. Best results are in bold , and second-best results are underlined .

Panel of Figure 6 — Figure 6: Boundary annotation refinement.

Stage	Criterion	Purpose
Boundary cue extraction	Boundary scale, regularity, interior consistency, and cross-boundary contrast	Retain closed structures that form coherent visible regions and provide clear perimeter-following cues.
Landmark cue extraction	Coarse-to-fine visual localization and local-view verification	Retain landmarks that can be localized from satellite crops and associated with nearby decision poses.
Route graph construction	Topological validity and visual observability of linear structures	Retain road and waterway segments that are sufficiently visible and can support unambiguous route-following instructions.
Trajectory instantiation	Visibility, smoothness, and action-space executability	Remove trajectories whose key cues are outside the local observation, whose turns are ambiguous or visually weak, or whose geometry cannot be followed by the discrete actions.
Instruction construction	Cue-trajectory consistency and language validation	Ensure that generated instructions refer only to cues that appear along the executable trajectory and remain semantically consistent after paraphrasing.

Appendix

Table 6: Automatic quality-control criteria used during SatNav episode generation.

Way Type	Degree	Local Geometry	Node Category
Highway	1	–	Road dead end
Highway	2	–	Waypoint
Highway	3	T-like	T-shaped intersection
Highway	3	Y-like	Y-shaped intersection
Highway	4	X-like	Crossroad
Waterway	1	–	River dead end

Appendix

Table 7: Mapping from local graph structure to node categories in the road and waterway graphs.

Panel of Figure 7 — Figure 7: Representative node types in the road and waterway graphs.

Figure 8: SatNav episode-quality review interface

Task	Subtype	Audited	Accepted	Rejected	Rejection (%)
Boundary	Partial-arc	155	144	11	7.1
	Full-loop	686	650	36	5.2
	Extended-loop	159	147	12	7.5
	Total	1,000	941	59	5.9
Landmark	One-turn	813	741	72	8.9
	Two-turn	187	170	17	9.1

Appendix

Table 8: Manual quality audit of 3,000 candidate episodes. Counts and rejection rates are reported by task family and subtype. All rejected episodes are removed before release.

Panel of Figure 9 — Figure 9: Examples of map memory. Blue indicates the start point, yellow indicates the current position and heading, and red indicates the trajectory. In each pair, the left is the global map and the right is the local map.

Figure 10: Data distribution and examples of the paired UAV-satellite images.

Method	Visual Input per Step	Action Output Horizon	Action Format
Seq2Seq & CMA	1 current	1 step	Categorical Logits
NaVILA	1 current + 7 history	1 step	Single-word text action
UniNaVid	1 current + all history	4 steps	Single-word text action
StreamVLN	current window + 8 history	4 steps	Symbolic (e.g., $\leftarrow,\uparrow$ )
OpenFly-Agent	1 current + 2 history + 16 past actions	1 step	Single-word text action

Appendix

Table 9: Summary of baseline adaptations for SatNav

Configuration	Seq2Seq	CMA
Vision Encoder	ResNet50	ResNet50
Instruction Encoding	GloVe-50d + GRU	GloVe-50d + BiGRU
Global Batch Size	64	32
Optimizer	Adam	Adam
Weight Decay	0.0	0.0
Peak LR	3e-4	1e-4

Appendix

Table 10: Training configurations for classical baselines on SatNav

Configuration	NaVILA	StreamVLN	UniNaVid	OpenFly-Agent	SwiftVLN
Base Model	Llama-3- VILA1.5-8B	LLaVA-Video- 7B-Qwen2	vicuna-7b- v1.5	openvla-7b- prismatic	Qwen2.5-VL- 3B-Instruct
Vision Encoder	SigLIP	SigLIP	EVA-CLIP ViT	DINOv2 + SigLIP	Qwen2.5-VL ViT
LLM Backbone	LLaMA-3-8B	Qwen2-7B	Vicuna-7B	LLaMA-2-7B	Qwen2.5-3B
Precision	bf16	bf16	bf16	bf16	bf16
Fine-tuning Strategy	Full	Full	Frozen Vision	Full	Full
#GPUs	8	8	8	8	8

Appendix

Table 11: Training configurations for LVLM baselines on SatNav

Figure 11: Real-world operation pipeline.

Model	Boundary				Landmark				Route
Model	SR $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps	SR $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps	SR $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps
Seq2Seq	2.1	78.7	65.11	45.88	0.0	0.4	238.47	55.03	4.4	12.2	243.57	59.42
CMA	12.4	92.4	63.54	80.05	0.0	2.9	789.61	159.76	13.9	32.0	214.13	59.58
OpenFly	16.4	66.9	52.03	74.10	2.7	7.1	288.36	54.06	20.2	26.9	160.36	49.15
OpenFly ∗	27.0	71.5	41.47	69.20	6.6	9.9	311.11	59.18	29.3	33.2	137.69	45.82
NaVILA	20.8	46.4	42.85	62.00	8.1	9.3	128.96	31.67	25.2	27.1	106.80	50.51

Appendix

Table 12: Per-task diagnostic results on Test Seen and Test Unseen. SR and OS are reported as percentages. Best results are in bold , and second-best results are underlined .

Base Model	Size	Seen					Unseen
Base Model	Size	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	NE $\downarrow$	Steps
Qwen2.5-VL	3B	65.8	65.5	72.4	34.05	49.97	53.7	53.2	64.1	62.29	57.38
Qwen2.5-VL	7B	70.2	69.8	77.4	29.97	48.18	56.6	56.2	66.2	55.33	54.62
Qwen3-VL	2B	67.5	66.7	79.5	52.64	58.13	57.0	56.3	70.3	71.94	65.84
Qwen3-VL	8B	68.3	67.6	76.7	45.12	52.16	56.0	55.3	64.9	69.79	58.55

Appendix

Table 13: Performance comparison of base vision-language backbones on SatNav. SR, SPL, and OS are reported as percentages. Best results are in bold , and second-best results are underlined .

Method	Image perturbation	SR $\uparrow$	SPL $\uparrow$	OS $\uparrow$	NE $\downarrow$
SwiftVLN	None	34.0	33.5	47.9	103.26
SwiftVLN w/ adapter	None	43.6	42.3	59.6	62.33
SwiftVLN	Exposure + Jitter	28.7	27.8	44.7	121.02
SwiftVLN w/ adapter	Exposure + Jitter	39.4	38.1	56.4	70.64

Appendix

Table 14: Navigation performance on 94 episodes rendered from real-world 3D assets. SR, SPL, and OS are reported as percentages, and NE in meters.

Factor	Setting	Train time	Eval time	Seen SR	Unseen SR
Baseline	Default	4:19:45	2:24:12	65.8	53.7
Satellite resolution	Zoom Level 17	4:24:19	2:29:25	64.9	52.8
	Zoom Level 15	4:12:06	2:23:27	62.4	52.5
Crop size	$336\times 336$ px	2:37:35	2:37:06	65.0	52.0
	$224\times 224$ px	1:33:34	2:50:05	51.1	40.8
Spatial coverage	$50\times 50$ m	4:16:34	3:02:32	56.4	46.3

Appendix

Table 15: Sensitivity of SwiftVLN to benchmark configuration. Each variant changes one factor from the default setting. Best SR results are in bold , and second-best results are underlined .

Panel of Figure 12 — Figure 12: Visualizations of trajectory examples for the Boundary task family.

Panel of Figure 13 — Figure 13: Visualizations of trajectory examples for the Landmark task family.

Panel of Figure 14 — Figure 14: Visualizations of trajectory examples for the Route task family.

Panel of Figure 15 — Figure 15: Visualizations of real-world UAV navigation examples, including a Route task and a Landmark task.

Figure 16: Representative real-world failure case on a Boundary task. The UAV is instructed to follow the boundary of the lake and stop after one complete loop, but fails to issue stop near the goal and continues circling for about two and a half loops.

VLN Benchmarks

See all 14 papers

Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation
Jun 5, 2026Xiangyi Zheng, Xiangyu Wang, Qinan Liao +6UAV Navigation Vision-Language Navigation
Colab, Beihang University · National University of Singapore · Meituan
VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method
Jul 23, 2026Jiabin Lou, Haopeng Wang, Yuanshuai Wang +6Embodied Navigation Robot Navigation
Beihang University, Beijing 100191, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China
FineCog-Nav: Integrating Fine-grained Cognitive Modules for Zero-shot Multimodal UAV Navigation
Apr 17, 2026Dian Shao, Zhengzheng Xu, Peiyang Wang +4Zero-Shot Learning UAV Navigation
Northwestern Polytechnical University · Nanjing University

Jun 5, 2026J/K move · Enter open · S save

Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation

Xiangyi Zheng, Xiangyu Wang, Qinan Liao, Zimu Tang+5

Colab, Beihang University · National University of Singapore · Meituan

UAV Navigation Vision-Language Navigation VLM Reasoning Vision-Language-Action Models Aerial VLN

Open paper AskPDF arXiv

Language-guided UAV agents must execute long-horizon semantic instructions while producing smooth, physically feasible continuous flight commands, yet existing Vision-Language Navigation (VLN) benchmarks typically use discrete or coarse actions and existing UAV Vision-Language-Action (VLA) tasks focus on short, atomic maneuvers. To address this gap in UAV task settings, we introduce FLIGHT, a Fine-grained Long-horizon Instruction-Guided benchmark for Hybrid UAV navigation and reasoning Tasks, which combines multi-stage instructions with dense 6-DoF trajectory annotations across two dataset splits: Fine-grained VLN and Long-horizon Flow. To endow the UAV agent with the capability of real-time in-flight reasoning over task execution status and mission planning, while simultaneously accommodating high-frequency, real-time precise control, we further propose FLIGHT VLA, an asynchronous architecture that decouples a low-frequency Streaming Pilot Vision-Language Model (VLM) for task-state reasoning from a high-frequency diffusion action model for continuous control, supervised by explicit Pilot Reasoning texts that summarize the current flight state and anticipate the next subgoal. In closed-loop evaluation, FLIGHT VLA consistently surpasses representative VLN and VLA baselines on our FLIGHT benchmarks, achieving stronger multi-stage completion, subgoal adherence, and terminal control. Its trained Streaming Pilot Reasoning VLM further improves UAV video reasoning, validating the effectiveness of our design.

Abstract

VLN Benchmarks

Explore similar work

Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation

VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method

FineCog-Nav: Integrating Fine-grained Cognitive Modules for Zero-shot Multimodal UAV Navigation