摘要 : Reinforcement learning (RL) methods are commonly regarded as effective solutions for designing intelligent driving policies. Nonetheless, even if the RL policy is converged after training, it is notoriously difficult to ensure saf... 展开 Reinforcement learning (RL) methods are commonly regarded as effective solutions for designing intelligent driving policies. Nonetheless, even if the RL policy is converged after training, it is notoriously difficult to ensure safety. In particular, RL policy is susceptible to insecurity in the presence of long-tail or unseen traffic scenarios, i.e., out-of-distribution test data. Therefore, the design of the RL-based decision-making method must account for this shift in distribution. This paper proposes a robust decision-making framework for autonomous driving on the highway to improve driving safety. First, a Deep Deterministic Policy Gradient (DDPG)-based RL policy that directly maps observations to actions is constructed. Subsequently, the model uncertainty of the DDPG policy is evaluated at runtime to quantify the policy's reliability and identify unseen scenarios. In addition, a complementary principle-based policy is developed using the intelligent driver model (IDM) and the model for minimizing overall braking induced by lane changes (MOBIL). It will take over the DDPG policy when encountering unseen scenarios to guarantee a lower-bound performance of the decision-making system. Finally, the proposed method is implemented on an embedded system, i.e., NVIDIA Jetson AGX Xavier, and out-of-training distribution challenging cases are considered in the experiment, i.e., observation with sensor noise, traffic density increasing significantly, objects falling from the front vehicle, and road construction causing temporal changes in road structure. Results indicate that the proposed framework outperforms state-of-the-art benchmarks. Additionally, the code is provided. 收起