SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery
SatNav is a satellite-imagery UAV vision-language navigation benchmark with 118,000 episodes across 18 cities.
SatNav is a scalable long-horizon UAV vision-language navigation benchmark built from high-resolution satellite imagery instead of costly 3D reconstructions. It contains 118,000 episodes from 59 scenes across 18 cities, with average trajectories of 379 meters, and three task families: Boundary, Landmark, and Route. Classical VLN agents and large vision-language model agents struggle at city scale. The authors also present SwiftVLN, a modular framework for memory ablations, and show satellite-trained models can transfer to real-flight UAV observations.
- 118K episodes from 59 scenes across 18 cities
- Average trajectory length is 379 meters
- Tasks cover Boundary, Landmark, and Route navigation
- SwiftVLN ablates switchable memory components
- Satellite-trained models transfer to real UAV flights
From the original article
1 min read at huggingface.co
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.
By Jiajun Jiang, Chunliang Hua, Zichun Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.