ForeFly: A Dual-Horizon World Action Model for Aerial Vision-Language Navigation

2026-09-27

AI Summary

This study addresses the unreliable instruction following of UAVs during long-horizon navigation in complex 3D environments, a limitation arising from the absence of multi-timescale lookahead cues. To this end, we propose a Dual-Horizon Latent World-Action Model that jointly predicts proximal local continuity and distal path-critical futures. We introduce a novel lookahead-guided action refinement mechanism that asymmetrically leverages dual-horizon predictions to enhance local action precision and global path correction, respectively. By integrating visual memory querying with feature-level action refinement, the proposed model significantly outperforms strong baselines across both seen and unseen scenarios on the TravelUAV and UAV-ON benchmarks. These results validate the effectiveness of dual-horizon lookahead modeling for robust UAV instruction-following navigation.

Abstract

Aerial Vision-Language Navigation (AVLN) requires UAVs to maintain reliable instruction following over long trajectories in complex 3D environments. However, existing AVLN approaches are predominantly reactive or limited to single-horizon prediction, overlooking complementary future cues across different temporal horizons. To address this limitation, we propose ForeFly, a dual-horizon latent world action model that predicts both a proximal future for local continuity and an adaptive route-critical future for long-range guidance. Horizon-specific foresight queries are primed with recent and route-critical visual memories, providing history-aware context for future prediction. To exploit their distinct roles in action generation, we introduce Foresight-Guided Action Refinement (FGAR), which asymmetrically exploits proximal foresight for local action enhancement and route-critical foresight for feature-wise correction and route-level guidance. Experiments on the TravelUAV and UAV-ON benchmarks show that ForeFly consistently outperforms strong baselines across seen and unseen settings, validating the effectiveness of dual-horizon foresight and FGAR learning. The code is available at: https://github.com/kunhuiW/ForeFly

Problem

• Aerial Vision-Language Navigation
• UAV
• long-horizon prediction
• 3D environments
• instruction following

Innovation

• Aerial Vision-Language Navigation
• Dual-Horizon World Model
• Latent Action Prediction
• Foresight-Guided Action Refinement
• Visual Memory