Research map / Vision-language and multimodal / Robot action and navigation
SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SatNav, a satellite-imagery-based long-horizon UAV vision-language navigation benchmark with 118K auto-generated episodes over 18 cities, plus the SwiftVLN modular LVLM navigation framework and memory ablations.
Problem
Scaling UAV vision-language navigation benchmarks is costly because they rely on reconstructed 3D assets, and existing benchmarks poorly test long-horizon memory and geospatial reasoning over city-scale routes.
Approach
Use high-resolution satellite crops (rotate-and-crop) as approximations of UAV nadir views, with episodes auto-generated from satellite imagery and OSM annotations across three task families (Boundary, Landmark, Route), and a modular SwiftVLN framework that swaps memory components for controlled ablations.
Key results
- StreamVLN fine-tuned on SatNav (StreamVLN*) reaches 58.4% SR on Test Unseen; SwiftVLN reaches 53.7% SR and the lowest NE (62.29 m) on Test Unseen.
- Temporal-biased history sampling and segment token clustering give small gains; removing long-term memory drops Test Unseen SR from 53.7 to 32.4.
- Real-world DJI flights (18 episodes): SwiftVLN with visual adapter raises overall SR from 38.9% to 55.6%.
Surprising findings
- OpenFly fine-tuned on SatNav substantially underperforms other LVLM baselines despite a UAV-pretrained initialization.
- Map memory and global token clustering underperform the reference model, while adding overlapping short-term windows gives the largest gains.
Limitations
- Fixed altitude and four discrete actions limit trajectory coverage and UAV mission types.
- No reinforcement fine-tuning; training only on reference trajectories.
- Real-world evaluation is limited to 18 episodes with a simple visual adapter.
Experimental setup
Datasets: SatNav cross-view geo-localization datasets (24K UAV-satellite pairs)
Benchmarks: SatNav Test Seen SatNav Test Unseen
Metrics: SR SPL OS NE Steps
Baselines: Seq2Seq CMA OpenFly NaVILA UniNaVid StreamVLN
Builds on: StreamVLN UniNaVid NaVILA OpenFly ms-swift
Closest papers by meaning
- Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology (ICLR 2025 · similarity 0.91)
- Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation (CVPR 2026 · similarity 0.89)
- WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.89)
- Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method (CVPR 2025 · similarity 0.88)
- Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory (CVPR 2026 · similarity 0.88)
- Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.