Certifying Learning-Enabled Autonomous Cyber Physical Systems - A Deployment Perspective

Deployment of Autonomous Cyber-Physical Systems (CPS) has penetrated almost every realm of human existence, including public transportation, military applications, critical infrastructures, and healthcare. While learning-enabled control methodologies for such systems promise scalability for massively distributed (yet coordinated) implementations, certification of safety and robustness is still considered elusive in that context. This is primarily because current strategies mostly rely on certification methods that are not scalable and inefficient in handling non-linearity. Such systems’ model-based control design and validation methods rarely account for the deployment platform uncertainties and adversarial aspects leading to attack exposed solutions. This work comprehensively summarises existing methodologies and critical issues related to real-world deployment of multi-agent learning-enabled CPS and also provides potentials solutions that explore innovative research roadmaps.