| Large vision-language models (VLMs) possess strong zero-shot semantic understanding capabilities, offering a promising approach to reducing the dependence of UAV vision-language navigation on task-specific training data. However, directly employing VLMs to predict navigation goals from urban overhead maps presents two major challenges. First, UAV headings are represented in the world coordinate system, whereas visual observations are defined in the image coordinate system. Existing methods lack explicit modeling of the transformation between these two coordinate systems, which can lead to errors in directional understanding. Second, navigators typically predict target pixel coordinates directly, making them susceptible to out-of-bounds predictions and localization errors despite correct semantic matching. To address these issues, we propose a zero-shot dual-agent collaborative navigation method for the CityNav task. The method decomposes high-level navigation reasoning into a perception agent and a planning agent. Through direction-aware projection, the perception agent maps the UAV heading from the world coordinate system to the image coordinate system, thereby establishing a unified coordinate reference for directional reasoning. Building upon this projection, the agent combines local-region refinement with candidate-node management to organize target-relevant visual evidence, spatial relationships, and landmark information into a structured candidate set. The planning agent then integrates multi-source evidence to perform constrained discrete selection over the candidate set, reformulating open-ended coordinate generation as candidate-node selection. The selected node is converted into world coordinates through a deterministic coordinate transformation, based on which a local controller generates UAV actions. The entire navigation process requires no task-specific fine-tuning. Experimental results on the CityNav dataset show that the proposed method achieves a success rate of 22.03% and an oracle success rate of 44.16% on the unseen test split, outperforming the state-of-the-art method by 0.83 and 8.78 percentage points, respectively. These results demonstrate the effectiveness of the proposed LVLM-based dual-agent collaborative navigation method. |
