Safety Assessment of a Machine Learning-Based Runway Detector

Xwing proposes to develop a black-box verification framework to facilitate the certification of safety-critical machine learning subsystems. Xwing, a leader in the development of autonomous flight technology, is investigating the safety assessment of machine learning models. Machine learning has demonstrated impressive results in sensing tasks relevant to autonomous flight. One such example is improving the safety and performance of runway detection in camera images. However machine learning cannot be substantiated following the standard DO-178 framework for software. Xwing thus proposes to investigate how to implement IASMS for the safety assessment of machine learning based subsystems. By leveraging an in-house simulator and runway detector, Xwing will investigate efficient sampling methods to estimate the probability of failure of the subsystem. The proposal focuses on three of the IASMS pillars: assessing the safety of the current model, mitigating nbsp;the failures modes exhibited through the framework, and assuring a revised model for certification. The framework will be validated with flight test data and used to produce example certification artifacts.