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

Traditional Vision-Language Navigation (VLN) for UAVs often treats navigation as a holistic problem, combining long-range search with final target approach. This paper argues that this approach masks failures in terminal precision. The authors ask: how can we isolate and evaluate the specific capability of a UAV to ground a visible target and execute precise 3D motion to reach it?

To address this, the authors formalize a new task called UAV-VLN-FOV, which focuses solely on the 'see-and-reach' stage. They introduce a high-resolution benchmark containing 2,717 trajectories with continuous 3D waypoint annotations. To solve this task, they propose 3DG-VLN, a framework that:

• Uses high-resolution front-view and downward-view observations to capture fine-grained visual and geometric details.
• Employs an online 3D direction updating strategy to re-estimate the target's relative position during flight, mitigating the spatial drift common in static-cue models.
• Utilizes a vision-language waypoint prediction model (based on Qwen2.5-VL) to translate instructions and visual cues into smooth, executable 3D trajectories.

The 3DG-VLN framework significantly outperforms existing UAV-VLN baselines in the see-and-reach task, achieving a 13.82% improvement in success rate. By mandating a stringent 10-meter success radius, the authors demonstrate that their model maintains superior terminal control fidelity compared to models designed for broader, less precise search tasks. Real-world trials confirm that the framework is robust enough for practical deployment in scenarios requiring high-precision navigation.