Assured Integration of Machine Learning-based Autonomy on Aviation Platforms

Dynamic assurance cases (DACs) are a novel concept for the provision of assurance-both during development and, subsequently, continuously in operation-that can be usefully applied to machine learning (ML)-based autonomous systems. We describe the application of a DAC for dependability assurance of an aviation system that integrates ML-based perception to provide an autonomous taxiing capability. Specifically, we present how we: i) formulate and capture risk-based safety and performance objectives, ii) model architectural mechanisms for risk reduction, iii) record the rationale that justifies relying upon autonomy, itself underpinned by heterogeneous items of verification and validation evidence, and iv) develop and integrate a computable notion of confidence that enables a run-time risk assessment and, in turn, dynamic assurance. We also describe our evaluation efforts, currently based on a hardware-in-the-loop simulator surrogate of an airworthy flight platform.