Neural Inertial Navigation Method for Wheeled Robots Based on Semi-supervised Co-training

Traditional supervised neural inertial navigation methods for wheeled robots rely much on training data. The lack of labeled data is usually caused by factors such as cost, which affects the normal training process This paper presents a semi-supervised one. It denoises IMU gyro data for attitude and faster iteration, estimates speed via co-training, and calculates the robot's position. Verified on a public dataset, without enough labeled data, compared to supervised methods, it reduces vehicle's absolute and relative trajectory errors by 83.27% and 84.14% respectively, proving its efficacy.