Deep Reinforcement Learning for Collision Avoidance of Autonomous Vehicle

Abstract

To save training efforts, reinforcement learning approach is applied to the autonomous vehicle for obstacle avoidance. Therefore, this study is aimed to let the autonomous vehicle to learn from mistakes and readdress its movement accuracy for collision avoidance in working environment. An enhanced learning method Q-learning is used to record and update the Q values for different movement through a table that the autonomous vehicle can use it to determine how and where to move. The Q table is learned through the deep learning neural network which may encounter innumerable situations from the environments and the different actions performed by the autonomous vehicle. In the experiments, the depth camera is adopted as the input device to be not affected by light intensity and road color. The Q table is ready to use after 9000 epochs or about 3.5 hours training. Let the autonomous vehicle run for 3 minutes at a time in three different environments with lights on and off 10 times each. The success rate of obstacle avoidance is as high as 95% which proves the feasibility of proposed approach.