Autonomous Motion Decision-making based on Deep Reinforcement Learning for Autonomous Driving

Motion decision-making is an open-challenging issue for autonomous driving, especially in the complex and diverse environment. A motion decision-making model based on deep reinforcement learning (DRL) is proposed in this work. To optimize the driving policy, a multi-objective reward function is designed to guide the autonomous driving system to explore optimal decision policy with the goal of safety, efficiency, and smoothness. A convolutional neural network (CNN) is designed as the backbone network of the DRL model to make full use of the observation information of environment. To enhance safety, a safety check module is designed for avoiding potential unsafe actions. A series of contrast experiments are conducted to justify the model performance of the proposed method on the DeepTraffic simulation environment. The experimental results demonstrate the effectiveness of the proposed DRL-based model in terms of the safe, efficient and smooth motion decision-making, and the generalization ability in different traffic conditions.