Testing Learning-Enabled Cyber-Physical Systems with Large-Language Models: A Formal Approach

The integration of machine learning (ML) into cyber-physical systems (CPS)offers significant benefits, including enhanced efficiency, predictivecapabilities, real-time responsiveness, and the enabling of autonomousoperations. This convergence has accelerated the development and deployment ofa range of real-world applications, such as autonomous vehicles, deliverydrones, service robots, and telemedicine procedures. However, the softwaredevelopment life cycle (SDLC) for AI-infused CPS diverges significantly fromtraditional approaches, featuring data and learning as two critical components.Existing verification and validation techniques are often inadequate for thesenew paradigms. In this study, we pinpoint the main challenges in ensuringformal safety for learningenabled CPS.We begin by examining testing as the mostpragmatic method for verification and validation, summarizing the currentstate-of-the-art methodologies. Recognizing the limitations in current testingapproaches to provide formal safety guarantees, we propose a roadmap totransition from foundational probabilistic testing to a more rigorous approachcapable of delivering formal assurance.