Future risk evolution is critical for the decision-making of autonomous vehicles (AVs), since it enables AVs to be aware of risks in advance and take action to ensure safety. This study proposes a reinforcement learning-based decision-making method by fusing future speculated risk for mixed traffic. Based on the current state of the environment and the predicted trajectories of pedestrians and vehicles, a driving risk field fusing current and future speculated risk is constructed to describe the driving risk. A Twin Delayed Deep Deterministic Policy Gradients (TD3) reinforcement learning algorithm is introduced as the main framework for decision-making training. The reward function accounting for both risk and efficiency is included in the TD3 training framework to optimize safety and efficiency performance. The proposed method is evaluated with the CARLA simulator in two scenarios (forward and unsignalized intersection scenarios). The experimental results show that the proposed method achieves better driving performance compared to methods that only consider instantaneous risk and rules.
Cai et al. (Fri,) studied this question.