Generalized Visual Path Following on Jetbot Using Normalization with Reinforcement Learning

摘要

Path following in robotics is pivotal for enabling autonomous navigation. However, its efficacy is often hindered by challenges such as varying environments, diverse weather conditions, and unknown parameters, which limit its adaptability. This study presents a practical application of the Reinforcement Learning Algorithm on physical robots for visual path-following tasks. It introduces a novel approach aimed at expediting training and enhancing generalization using normalization techniques. Our experimentation entails applying Deep Q learning to JetBot—a nonholonomic wheeled mobile robot—to assess the generalization capabilities of normalization techniques including Batch Normalization, Layer Normalization, Min-Max Normalization, and Z-score Normalization. Our findings reveal that the model proposed in this study, incorporating Batch Normalization, achieves 8.7 times higher expected rewards during training, operates 5 times faster, and yields 3 to 4 times greater expected rewards during testing compared to the simple Deep Q learning model.