Skip to content
Book Open access

UAV Vision-Language Navigation via Instruction-History Interactive Filtering and Mamba-based Temporal Modeling

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 26 references

Abstract

Unmanned aerial vehicle (UAV) vision-language navigation requires agents to perform autonomous navigation in complex 3D environments by following natural language instructions. Existing methods face critical challenges when processing long sequences of historical observations, including visual redundancy and insufficient temporal context modeling. To address these issues, we propose a UAV vision-language navigation framework that integrates instruction-history interactive filtering with Mamba-based temporal modeling. First, we design an Instruction-Guided Keyframe Selection (IGKS) mechanism that dynamically extracts semantically relevant keyframes based on interaction between the instruction and historical observations, effectively compressing input sequences while preserving spatiotemporal context. Subsequently, we introduce the state space model Mamba to efficiently model temporal dependencies among the selected keyframes, capturing implicit motion trends and environmental evolution patterns during navigation, thereby enhancing the agent's understanding of historical trajectories. Finally, the resulting temporal features are used to predict the next action with a large language model. Experimental results on public UAV VLN dataset demonstrate its effectiveness in terms of navigation success rate and path efficiency, particularly exhibiting stronger robustness and semantic alignment in complex scenarios.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.