Accurate odometry estimation is critical for the reliable operation of intelligent systems. Despite the remarkable progress of deep learning–based odometry models in controlled environments, real-world degradations such as IMU noise, visual impairments, and calibration errors often lead to significant estimation failures. Conventional testing methods, constrained by limited datasets and costly manual labeling, are insufficient for systematically revealing these vulnerabilities. In this paper, we propose OdoTest, the first automated testing framework for odometry models. OdoTest leverages odometry-specific metamorphic relations and introduces transformation operators that simulate realistic sensor and calibration perturbations, enabling the generation of transformed test datasets that expose model weaknesses. An odometry-oriented scenario generation strategy further improves testing efficiency. We evaluate OdoTest across multiple odometry models, and the results demonstrate that it can effectively detect erroneous behaviors under diverse odometry-specific conditions. Additionally, retraining with the generated scenarios improves odometry precision and reduces estimation errors, highlighting OdoTest’s potential for enhancing model reliability.