Abstract
UAV see-and-reach navigation requires an aerial agent to approach a language-specified target visible in its initial view and stop reliably near it. Existing UAV-VLN methods either directly map vision-language representations to actions or employ intermediate reasoning mainly for semantic guidance, while control-oriented spatial decisions for maneuver and termination remain implicit. This limits the alignment between semantic understanding and continuous UAV motion, leading to inconsistent maneuvers and unreliable termination. To address this issue, we propose DBFly, which introduces control-oriented spatial decision decomposition into UAV see-and-reach navigation to explicitly structure maneuver and termination decisions before continuous waypoint generation. Specifically, DBFly introduces a spatial maneuver decision chain that progressively performs target-direction anchoring, spatial diagnosis, and maneuver decision, enabling explicit maneuver decisions to condition continuous waypoint generation. DBFly further constructs an implicit flight corridor by transforming the initial target-direction prior into a persistent geometric reference and deriving an online corridor state from the UAV's current position, thereby providing soft geometric guidance for spatial diagnosis and maneuver correction. In addition, DBFly develops a terminal-convergence-aware stopping strategy that characterizes terminal states through both target proximity and short-horizon motion convergence, enabling more reliable stopping near the target. Extensive experiments across seen, unseen-object, and unseen-scene test sets demonstrate that DBFly improves the success rate over the SOTA baseline by an average of 18.57 percentage points. Decision-level evaluations further demonstrate highly consistent alignment between maneuver decisions and generated waypoints, together with substantially improved stopping reliability.
Overview
Overall framework of DBFly.
Simulation Experiments
Representative closed-loop flight demonstrations in high-fidelity simulation environments.
Simulation Demo 2
Simulation Demo 3
Real-World UAV Experiments
Representative closed-loop flight demonstrations in real-world environments.
Real-World Demo 1
Instruction: Fly to the green ball on the yellow bike.
Real-World Demo 2
Instruction: Fly to the blue box in the woods.
Results Highlights
DBFly achieves consistent improvements across seen, unseen-object, and unseen-scene test settings.
BibTeX
Citation information will be updated after the paper is publicly available.
@misc{xue2026dbfly,
title={Deliberate Before You Fly: Vision-Guided Spatial Deliberation for UAV See-and-Reach Navigation},
author={Fanfu Xue and En Yu and Bohang Liu and Hongjun Wang and Yang Yang and Xindi Wang and Jiande Sun},
year={2026},
eprint={2608.04825},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2608.04825},
}