See-and-Reach: Precise Vision-Language Navigation for UAVs within the Field of View

Researchers introduce UAV-VLN-FOV, a new evaluation framework for unmanned aerial vehicle vision-language navigation that focuses on precise target reaching once the target is visible. The accompanying 3DG-VLN model uses dual-view observations and dynamic 3D direction cues to improve navigation accuracy by 13.82%, with real-world validation demonstrating practical viability.

This research addresses a fundamental gap in how autonomous aerial navigation systems are evaluated and trained. Traditional UAV vision-language navigation benchmarks combine long-range target discovery with final approach in a single holistic task, obscuring whether agents can accurately translate visual and linguistic cues into precise 3D movements once targets enter their field of view. By isolating the see-and-reach stage, this work enables more granular assessment of a critical capability that separates theoretical navigation from practical deployment.

The technical innovation centers on processing multiple camera perspectives simultaneously—front-facing and downward-looking views—to capture both fine-grained visual details necessary for grounding targets and geometric information crucial for spatial reasoning. The dynamic 3D direction updates during closed-loop navigation represent a practical solution to accumulated directional drift, a persistent challenge in aerial systems. The construction of a dedicated benchmark with 2,717 trajectories and continuous 3D waypoint annotations provides valuable infrastructure for the research community.

From an industry perspective, this work has implications for autonomous delivery systems, inspection drones, and search-and-rescue operations that increasingly rely on natural language instructions. The 13.82% improvement in success rates demonstrates measurable progress toward real-world deployment standards. Real-world trials validating the approach suggest these methods are approaching practical reliability thresholds. The open-sourcing of code and benchmark accelerates adoption and further research, potentially spurring development of commercial applications requiring precise visual-linguistic navigation in constrained environments.

• →UAV-VLN-FOV task isolates terminal reaching ability from long-range search, enabling more diagnostic evaluation
• →3DG-VLN framework achieves 13.82% success rate improvement through dual-view processing and dynamic direction updates
• →Real-world validation demonstrates practical viability beyond simulation environments
• →Open-sourced benchmark with 2,717 trajectories provides foundation for advancing autonomous aerial navigation
• →Multi-view sensor fusion combined with closed-loop directional correction addresses accumulated drift in navigation