Publication Type

PhD Dissertation

Version

publishedVersion

Publication Date

8-2026

Abstract

Autonomous Vehicles (AVs) rely on Autonomous Driving Systems (ADSs) to perceive their surroundings, predict the behavior of traffic participants, plan driving actions, and control vehicle motion in complex traffic environments. As AVs move toward large-scale real-world deployment, ADSs are expected not only to avoid collisions and traffic-rule violations, but also to make robust decisions, cooperate effectively with other AVs, and maintain high-quality driving behavior. However, the open, dynamic, and uncertain nature of driving environments makes the systematic testing and enhancement of ADSs a challenging problem.

Although scenario-based testing has made substantial progress in discovering safety-critical violations, several important limitations remain. Existing methods often prioritize the number of detected violations and may repeatedly expose similar failures, providing limited insight into the range of ADS weaknesses. They also focus mainly on safety-critical requirements while paying less attention to decision robustness and other non-safety-critical properties. Moreover, most techniques evaluate ADSs in single-AV settings, although interaction-level failures may emerge when multiple ADS-equipped vehicles make interdependent decisions. Finally, testing can reveal failures but cannot directly mitigate safety risks during runtime. This dissertation investigates these limitations from four connected perspectives: violation diversity, decision-making robustness, multi-AV cooperation, and adaptive runtime enhancement.

The first study proposes BehAVExplor, a behavior-diversity-guided testing technique for discovering diverse safety-critical failures. BehAVExplor characterizes temporal ego-vehicle behavior patterns from driving traces and uses behavioral diversity, together with violation-oriented feedback, to guide scenario generation. By balancing behavior exploration and violation discovery, it reduces redundant failure generation and exposes safety-critical scenarios associated with a broader range of ego-vehicle behaviors and potential system weaknesses.

The second study proposes Decictor, an automated framework for evaluating the robustness of ADS path-planning decisions. Decictor generates non-invasive scenario perturbations that are designed not to affect the original optimal path and examines whether the ADS preserves its planning decision after perturbation. A consistency-based oracle identifies scenarios in which the ADS remains safe but selects a non-optimal path. This study extends ADS testing beyond binary safety outcomes to the evaluation of decision quality and robustness.

The third study proposes WaitWatch, a testing framework for evaluating cooperative behavior in multi-AV scenarios. WaitWatch formalizes deadlocks as persistent circular wait-for relations among ADS-controlled vehicles and introduces a black-box deadlock oracle for detecting them from runtime observations. It further combines spatio-temporal feedback with intersection-oriented mutation to guide scenario generation toward deadlock-prone interactions. This study extends ADS testing from individual-vehicle behavior to interaction-level and cooperative failures involving multiple AVs.

The fourth study explores ADReFT, an adaptive runtime decision-repair method that combines supervised training and reinforcement fine-tuning to mitigate safety-critical decisions while limiting unnecessary intervention.

Overall, this dissertation advances automated ADS testing from safety-critical failure discovery to decision robustness and multi-AV cooperation, while also exploring adaptive runtime enhancement. The proposed techniques provide practical support and empirical insights for developing safer, more robust, and more capable ADSs.

Keywords

Autonomous Driving Systems, Software Testing, Decision-Making Robustness, Behavioral Diversity, Multi-AV Deadlock, Runtime Enhancement

Degree Awarded

PhD in Computer Science

Discipline

Artificial Intelligence and Robotics | Software Engineering

Supervisor(s)

XIE, Xiaofei

First Page

1

Last Page

216

Publisher

Singapore Management University

City or Country

Singapore

Copyright Owner and License

Author

Download

Share

COinS