SI-safe-visor: Sandboxing (AI-based) unverified controllers in linear parameter varying systems

Safe-visor architecture provides a fault-tolerant control architecture for cyber-physical systems controlled by unverified high-performance controllers (HPC), such as AI-based controllers. Specifically, Safe-visor architecture deploys a Supervisor to identify and reject the unsafe control inputs provided by the HPC. Meanwhile, a Safety Advisor in the architecture is used to provide fallback control input to ensure overall safety of the system in case that the HPC is rejected. However, the design of Safe-visor architecture is typically model-based, whose implicit assumption is the stored fixed normal model accurately captures the underlying dynamics in consideration, which potentially hinders the feasibility of Safe-visor architecture in parameter varying systems. To address this issue, we propose a system identification augmented Safe-visor (SI-Safe-visor) architecture, in which a least-squares-based system dynamics identifier is adopted for online estimation of varying system parameters and updating of control actuator in the Safety Advisor, as well as the Supervisor. A case study for an autonomous vehicle is provided to numerically demonstrate the effectiveness of SI-Safe-visor.