UAV–VLM–UGV Task Reasoning Evaluation Dataset

This dataset supports the quantitative evaluation of VLM-based task reasoning for UAV–UGV cooperative autonomous operation. It contains four scene categories: beach, grassland, paved ground, and playground. The dataset contains 645 annotated images and 2,640 task-oriented ground-truth zones, including 1,618 target zones, 345 obstacle zones, and 677 risk zones. Polygon coordinates are normalized to the range [0, 999]. All beach images were collected by the authors. In the grassland, paved-ground, and playground subsets, images whose filenames end with “AI” were generated using artificial intelligence. The remaining images were selected from the UAVVaste dataset (DOI: 10.5281/zenodo.8214061). Only the source images from UAVVaste were reused. The original UAVVaste annotations were not used. All target, obstacle, and risk zone annotations included in this dataset were independently created for the present study.