Abstract: Learning-enabled cyber-physical systems (CPS), such as autonomous driving and UAVs, are increasingly deployed in safety-critical settings, yet their reliance on foundation models poses challenges for testing and assurance. This seminar presents recent advances in using foundation models and LLMs for safety testing, including specification mining from rules and accident data, and automated scenario generation for conformance testing. It then highlights a transition toward neuro-symbolic architectures via NeuroStrata, integrating neural perception with symbolic reasoning for interpretable, verifiable assurance. A UAV landing-site case study demonstrates symbolic world models, differential rule engines, and human-in-the-loop validation. Insights from Shonan Seminar 235 outline a roadmap for trustworthy CPS.

Chair: Paolo Tonella