Abstract
The validation of cyber-physical systems (CPS) is a critical challenge, particularly in safety-critical domains such as autonomous vehicles (AVs) and robotic control systems. Traditional validation techniques often struggle to efficiently explore the vast state and action spaces of these systems, limiting their ability to uncover failure scenarios and assess robustness under diverse conditions. Recent advancements in Reinforcement Learning (RL) have introduced new methodologies for system validation, leveraging RL’s ability to discover optimal or near-optimal policies while systematically covering a wide range of possible system states. This capability makes RL particularly useful for stress-testing CPS algorithms by exposing them to rare or extreme conditions that traditional testing approaches might overlook. In this paper, we present an RL-based validation approach designed to evaluate the robustness and compliance of a cyber-physical system’s motion planning algorithm in simulation. The framework is applied to an adaptive cruise control (ACC) system in an AV, demonstrating its effectiveness in identifying failure cases where the system deviates from its specified design requirements. Compared to conventional testing approaches, our method provides a more systematic and automated means of discovering failure-inducing scenarios, ensuring broader coverage of the operational space. The results illustrate how RL-driven validation can complement traditional CPS verification techniques by efficiently identifying edge cases and potential safety violations. While this study focuses on AV applications, the proposed RL-based framework has the potential to extend to a wide range of CPS domains, including robotics, industrial automation, and aerospace systems. By integrating RL into the validation pipeline, this research contributes to the development of more resilient, compliant, and robust CPS architectures capable of operating reliably in complex, dynamic environments.