AerialDojo-200K
收藏资源简介:
AerialDojo-200K是一个大规模的开源世界空中目标搜索基准套件,旨在评估空中智能体在大型非结构化3D环境中自主探索并找到由语义描述(SemanticOGS)或参考图像(ImageOGS)指定的目标物体的能力。该套件包含42个仿真场景,涵盖4个场景家族和21种场景类型,共提供205,732个任务实例,分为基础、标准和长视域三种设置。其规模是先前最大空中目标搜索基准的3倍场景数和18.7倍任务实例数。数据集由12名标注员耗时两个月精心标注,包含109个地标、2,099个目标物体以及2,099个物体锚点,并提供了总长4,115.313公里的无碰撞参考轨迹和63,177组专用于训练的多视角录制数据。数据格式、动作空间和评估协议统一,支持在21个分布内场景和21个分布外场景上进行评估。对9个多模态大语言模型(5个开源和4个闭源)的基线评估突显了构建通用空中智能体面临的挑战。数据集包含AerialENVS(UE/ProjectAirSim环境包)、SemanticOGS(语义目标搜索任务)、ImageOGS(图像目标搜索任务及参考图像)和TrajectoryDATA(每条地图/任务/分区合并的轨迹文件)。注意:分布外测试集(OOD_TESTS)未公开,需联系作者进行评估。
AerialDojo-200K is a large-scale open-world aerial target search benchmark suite designed to evaluate the ability of aerial agents to autonomously explore and find target objects specified by semantic descriptions (SemanticOGS) or reference images (ImageOGS) in large unstructured 3D environments. The suite contains 42 simulation scenes covering 4 scene families and 21 scene types, providing a total of 205,732 task instances divided into three settings: basic, standard, and long-horizon. Its scale is 3 times the number of scenes and 18.7 times the number of task instances of the previous largest aerial target search benchmark. The dataset was carefully annotated by 12 annotators over two months, containing 109 landmarks, 2,099 target objects, and 2,099 object anchors, and provides a total of 4,115.313 km of collision-free reference trajectories and 63,177 sets of multi-view recording data dedicated to training. The data format, action space, and evaluation protocol are unified, supporting evaluation on 21 in-distribution scenes and 21 out-of-distribution scenes. Baseline evaluations of 9 multimodal large language models (5 open-source and 4 closed-source) highlight the challenges of building general-purpose aerial agents. The dataset includes AerialENVS (UE/ProjectAirSim environment package), SemanticOGS (semantic target search tasks), ImageOGS (image target search tasks and reference images), and TrajectoryDATA (merged trajectory files for each map/task/partition). Note: The out-of-distribution test set (OOD_TESTS) is not publicly available and requires contacting the authors for evaluation.





