Certifying Machine Learning for Autonomous Flight

Jean-Gullaume Durand, Xwing ABSTRACT: Over the past decade, machine learning (ML) has demonstrated impressive results, often surpassing human capabilities in sensing and perception tasks relevant to autonomous flight. Xwing is an autonomous aircraft company that leverages these research breakthroughs to improve the performance and safety of SuperPilot, its autonomy stack. In this talk, we will first present where Xwing uses ML to leverage a safety advantage. We will then break down why these ML models do not fit into the traditional aerospace certification paradigm. To overcome this limitation, we will introduce Xwing’s methodology to certification: a framework centered around a statistical verifier for black-box subsystems. The framework is model-agnostic and tool-independent, making it adaptable to many use cases in the industry. We will demonstrate results on vision-based landing, a widespread application in autonomous flight.