CAAD: A High-Level Customizable-Agent Gym for Dependable Autonomous Driving

We present CAAD, a comprehensive framework designed to enhance the dependability of reinforcement learningbased autonomous driving (RLAD) systems. The proposed framework provides a high-level agent architecture that enables customizable agent modeling and behavior definition. It supports flexible scenario customization through domain randomization techniques, while further enriching scenario complexity by integrating external policies and state-of-theart training algorithms. Additionally, the framework offers practical utilities for systematic dependability analysis and evaluation of autonomous driving systems. We demonstrate the framework's capabilities through several experiments. Firstly, the scenarios from existing frameworks are implemented in the proposed framework to highlight the flexible definition of the task scenarios. Then, we validate the impact of atomic actions in agent training and design various scenarios to generate critical scenarios for external systems. Finally, we illustrate the process of improving the dependability in RLAD systems. The comprehensive experiments and results demonstrate that the proposed framework effectively enhances the dependability of RLAD systems through its integrated modeling, testing, and analysis capabilities.