Rat robots have great potential in rescue and search tasks because of their excellent motion ability. However, most of the current rat-robot systems relay on human guidance due to variable voluntary motor behaviour of rats, which limits their application. In this study, we developed a real-time system to detect a rat robot’s transient motion states, as the prerequisite for further study of automatic navigation. We built the detection model by using a wearable inertial sensor to capture acceleration and angular velocity data during the control of a rat robot. Various machine learning algorithms, including Decision Trees, Random Forests, Logistic Regression, and SupportVector Machines,were employed to performthe classification of motion states. This detection system was tested in manual navigation experiments, with detection accuracy achieving 96.70%. The sequence of transient motion states could be further used as a promising reference for offline behaviour analysis.