Reinforcement learning-based end-to-end self-driving research using variational autoEncoder

Abstract

There are two main ways of learning and operating an autonomous vehicle. One is between finding an optimization relationship by learning without having to design a module by designing perception, decision, and control algorithm by using a camera, LiDAR, GPS, and HD maps. Another approach is end-to-end driving. Currently, end-to-end learning is already showing remarkable performance in many fields, and some areas already surpass human’s performance. Thus, it is expected that the end-to-end driving method will learn the optimal driving skills and find an optimized combination for input values under the assumption that it can collect enough data, and that it will show performance beyond the modular method defined by humans. In this paper, variational autoencoder(VAE) will be introduced, which can compress high-dimensional information to low-dimension. We will propose interpretable end-to-end driving based on reinforcement learning in continuous space using VAE. In addition, we will verify the simulation to the real world algorithm performance in the real world. Finally, one episode was designed by dividing it into several smaller episodes to make the sparse rewards dense, and we will verify that the proposed algorithm is a technology that effectively guides the end-to-end-based autonomous vehicle to reach the final destination.