There are two main approaches to learning and operating autonomous vehicles. One method is to drive based on perception, decision, and control modules using cameras, lidar, GPS, and precision maps. The other one is end-to-end driving combining whole modules without human’s inductive bias. Currently, end-to-end learning methods are showing remarkable performance in many areas. Even, some areas are already performing beyond humans. It is expected that the end-to-end driving method will outperform modular approach which is constructed by human. This is because it will find an optimized combination between input values by learning through big data on the assumption that it can collect enough data. In this paper, we will utilize variational autoencoder(VAE) that compresses a high-dimensional image to a low-dimensional vector that is easy to handle. We develop an end-to-end autonomous driving algorithm based on reinforcement learning in a continuous space using a variable as a state value, and verify its performance in an actual road environment. |
