Risk-aware Deep Reinforcement Learning for Decision-making and Planning of Autonomous Vehicles

To improve the safety and efficiency of autonomous vehicles on the highway, a hierarchical framework combining deep reinforcement learning and risk assessment is proposed in this paper for lane change decision-making and planning. Firstly, a new risk assessment method is proposed to accurately indicate the driving risk. Then, the quantitative risk is used to serve as the observation of the deep double Q-network algorithm and to design the reward function. After obtaining the driving behavior, a quintic polynomial is used to plan trajectory with considering vehicle dynamics constraints, safety, efficiency and vehicle comfort. The proposed framework is compared with the traditional Double DQN algorithm with the input of motion state (speed, acceleration, heading angle). The training and test results in SUMO demonstrated that the proposed method outperformed the previous methods in training efficient and driving safety.