LongFly: UAV Navigation with Spatiotemporal Context

This paper addresses the challenge of long-horizon vision-and-language navigation (VLN) for UAVs, a critical area for applications like search and rescue. The core contribution is a framework, LongFly, designed to model spatiotemporal context effectively. The focus on distilling historical data and integrating it with current observations is a key innovation for improving accuracy and stability in complex environments.

• •Proposes LongFly, a framework for long-horizon UAV VLN.
• •Employs a history-aware spatiotemporal modeling strategy.
• •Includes modules for image compression, trajectory encoding, and multimodal integration.
• •Achieves significant performance improvements over existing baselines.