Reliable Reinforcement Learning for Decision-Making in Autonomous Driving

Abstract

Autonomous driving technology has made significant strides due to advancements in artificial intelligence, sensor technology, and computational power. However, deploying autonomous vehicles (AVs) in real-world scenarios remains challenging due to safety concerns, the need for generalizability across diverse environments, and the demand for interpretable decision-making processes. This dissertation addresses these challenges by developing reliable reinforcement learning (RL) algorithms tailored for autonomous driving decision-making, providing a comprehensive framework for creating robust RL-based solutions and bridging the gap towards safer and more efficient autonomous transportation systems.

First, we introduce a safe-RL-based solution that ensures safety during both the training and deployment phases. This is achieved by formulating the learning problem as a constrained optimization problem and applying a parallel training strategy to enhance training efficiency and the likelihood of achieving an optimal policy.

Second, we present meta-RL-based solutions designed to enhance the generalizability of policies. By incorporating safety into the exploration of prior policies, we ensure the safety of the policy before adapting to new tasks. Additionally, leveraging task interpolation and data augmentation improves the data efficiency of current meta-RL techniques while maintaining the same level of generalization performance.

Third, we propose an interpretable decision-making solution through an intention-aware decision-making approach that uses a hierarchical architecture to generate driving intentions and corresponding trajectories. This approach improves the interpretability of the decision-making process, facilitating better interaction with surrounding traffic participants and enhancing overall system performance.

The effectiveness of these contributions is demonstrated through a series of experiments in simulated environments and datasets, focusing on tasks such as lane-keeping, intersection crossing, and highway merging. Our results show significant improvements in safety, generalizability, and interpretability, bridging the gap between simulation-based RL approaches and real-world deployment.