The Internet of Robotic Things (IoRT) has experienced rapid growth and garnered increased attention in recent years. Applications (Apps) play a crucial role in IoRT, as they provide users with an intuitive interface to access and operate services. However, as user demands increase, Apps become more complex, leading to longer operation sequences and more vulnerabilities. The existing testing methods for Apps can be categorized into random, reinforcement learning, and AI-based approaches. AI-based methods offer a solution to the low coverage efficiency of random-based methods and the weak guidance of reinforcement learning-based methods. However, current AI-based methods have difficulty in capturing long-term dependencies, resulting in low coverage and less detected crashes when testing Apps with long operation sequences. To address the limitation, we propose InforTest, a novel AI-based method based on the Informer prediction model and the component tree structure. InforTest leverages Informer, which excels at extracting long-term dependencies from operation sequences, to generate human-like moves for testing Apps. To improve the efficiency of training and prediction, InforTest uses the component tree, a concise structure to represent primary data sources, i.e., screenshots. After training InforTest on the Rico dataset, our experiments with Apps in the IoRT scenario demonstrated its superiority over existing methodologies such as Monkey, Humanoid, MUBot, and Ape. Notably, InforTest achieved significant enhancements in coverage rates (increases of 67%, 34%, 19%, and 27%, respectively) and in crash detection capabilities (improvements of 175%, 81%, 71%, and 139%, respectively).