D-GARA
收藏资源简介:
D-GARA是由同济大学等机构联合开发的动态基准测试框架,专注于评估Android图形界面智能体在真实异常环境下的鲁棒性。该数据集通过集成Android模拟器构建动态交互环境,包含权限弹窗、系统警告等高频异常类型,支持实时异常注入与多路径执行轨迹模拟。其创建过程采用语义触发机制与可配置规则库,通过数据收集工具采集带有人工标注的屏幕截图与XML文件。该数据集主要应用于智能体鲁棒性评估领域,旨在解决现有静态基准无法反映真实环境动态复杂性的核心问题,推动强适应性GUI智能体的发展。
D-GARA is a dynamic benchmark testing framework co-developed by Tongji University and other institutions, dedicated to evaluating the robustness of Android GUI agents in real-world anomalous environments. This dataset establishes a dynamic interactive environment through the integration of Android emulators, incorporates high-frequency anomaly types including permission pop-ups and system warnings, and supports real-time anomaly injection and multi-path execution trajectory simulation. Its development process adopts a semantic triggering mechanism and a configurable rule base, and collects manually annotated screenshots and XML files via dedicated data collection tools. This dataset is primarily applied in the field of agent robustness evaluation, aiming to address the core limitation that existing static benchmarks fail to reflect the dynamic complexity of real-world environments, and advance the development of highly adaptive GUI agents.
D-GARA 数据集概述
数据集名称
D-GARA
核心功能
动态基准测试框架,用于评估图形用户界面(GUI)智能体在真实世界安卓异常情况下的鲁棒性。
项目信息
- 项目页面:https://sen0609.github.io/D-GARA/
- 代码仓库:https://github.com/sen0609/D-GARA
- 引用文献:Chen等人,AAAI 2026会议论文
应用领域
安卓GUI智能体测试、鲁棒性评估、动态基准测试

- 1D-GARA: A Dynamic Benchmarking Framework for GUI Agent Robustness in Real-World Anomalies同济大学 · 2025年



