Tequila510/Mobile-RobustBench
收藏资源简介:
Mobile-RobustBench是一个用于系统评估移动GUI代理在真实世界视觉扰动下鲁棒性的基准数据集。该数据集包含9个应用程序、143个任务和712个跟踪记录,平均每个跟踪记录有5.48个步骤。数据集结构包括数据文件夹(包含应用程序的截图、UI层次结构和标注文件)、评估文件夹(包含计算指标的脚本)等。数据格式详细描述了每个应用程序的JSON文件结构,包括任务描述、干净跟踪记录和扰动跟踪记录。此外,还详细列出了10种扰动类型及其影响。
Mobile-RobustBench is a benchmark for systematically evaluating the robustness of mobile GUI agents under real-world visual perturbations. The dataset includes 9 apps, 143 tasks, and 712 traces, with an average of 5.48 steps per trace. The dataset structure consists of a data folder (containing app screenshots, UI hierarchies, and annotation files), an evaluation folder (containing scripts for calculating metrics), etc. The data format details the JSON file structure for each app, including task descriptions, clean traces, and perturbed traces. Additionally, it lists 10 types of perturbations and their impacts.





