DroidGamer: Android Game Testing with Operable Widget Recognition by Deep Learning

Android game applications are an important type of application widely used by end users. Bugs in such applications can significantly affect user experience. Due to the use of rendered Graphical User Interface (GUI) widgets, automated testing of Android game application becomes challenging because such GUI widgets cannot be queried with Android system APIs, making existing GUI-based testing techniques blind to the locations of widgets. In this work, we propose DroidGamer, a novel GUI traversal-based Android game testing technique, which relies on deep learning models to recognize operable GUI widgets. DroidGamer adopts a novel GUI model traversal algorithm and a new GUI state equivalence criterion over the widget recognition results of the deep learning models. Our experiment on 10 open-source Android games shows that DroidGamer is significantly more effective than existing techniques including Monkey, Stoat, and PUMA for testing Android games in terms of both code coverage and fault detection ability.