Neural inertial navigation method for wheeled robots based on self-supervised learning

Low-cost Micro-Electro-Mechanical System Inertial Measurement Units (MEMS-IMUs) are plagued by large, complex, and variable errors. Traditional strap-down inertial navigation systems that utilize MEMS-IMUs are unable to meet the positioning requirements of wheeled robots. Although inertial navigation based on deep learning has been explored, it necessitates a substantial amount of carefully selected and labelled data, resulting in high costs. Consequently, this paper proposes a self-supervised neural inertial navigation method for wheeled robots that solely depends on MEMS-IMU data. Firstly, a representation learning model is established to extract general IMU features for self-supervised denoising. Subsequently, an intelligent framework employing contrastive learning is adopted to explore the latent information of the IMU and acquire the motion state of the robot. Specific motion state information is regarded as observations, and an invariant extended Kalman filter (IEKF) is applied for information fusion to enhance positioning accuracy. Experiments conducted on public datasets demonstrate that, in the absence of additional ground truth values, the Absolute Trajectory Error (ATE) and Temporal Relative Trajectory Error (T-RTE) of the proposed method are 20.23% and 30.71% lower than those of supervised learning-based methods, respectively. The proposed method offers a more cost-effective and practical solution for the development of inertial navigation technology for wheeled robots.