Towards Autonomous Driving Decision by Combining Self-attention and Deep Reinforcement Learning

Abstract

Autonomous driving decision-making is a challenging task for the complexity of the environment. The existing methods are rule-based and supervised learning methods, but these methods can only get suboptimal strategies. In recent years, using deep reinforcement learning(DRL) to complete autonomous driving decision-making task has gained widely attention. In this paper, we propose an algorithm framework based on self-attention model and DRL to solve the problem of vision-based autonomous driving decision in complex scenarios. We use the self-attention model to reduce the dimension of image and get global features. Then we use deep deterministic policy gradient(DDPG) algorithm to complete the autonomous driving decision-making task. We evaluate our method in the complex scenario provided by Carla. The results show that our method can learn better strategies with higher efficiency and reward. In addition, we also visualize the output of the self-attention model, and the results show that our model can identify the position of obstacles in the image, and improve the interpretability of the model.