ISBN 9783843957465

978-3-8439-5746-5, Reihe Informationstechnik

Christoph Regli
On the Training and Checking of AI Pilots – An Anthropomorphic Approach to Leverage Trust through Licencing

332 Seiten, Dissertation Universität Stuttgart (2026), Hardcover, B5

Zusammenfassung / Abstract

This dissertation addresses the challenge of training and checking an Artificial Intelligence (AI) pilot capable of operating under the inherently unpredictable conditions of flight operations. Human pilots rely on structured training and recurrent checking to develop problem-solving skills and the ability to transfer learning to unforeseen situations — capabilities that conventional rule-based systems cannot replicate due to the vast, dynamic input space of real-world aviation. As a result, future automated flight systems are expected to combine traditional software with Machine Learning (ML) components, enabling learning from operational data and generalization from past experience while maintaining robustness and safety.

The integration of ML into avionics introduces adaptability but potentially non-deterministic behaviour, rendering conventional validation, verification, and certification approaches insufficient. Current regulatory initiatives primarily follow a bottom-up approach, focusing on development processes and training data integrity. This thesis complements these efforts with a top-down, performance-based "anthropomorphic" approach, inspired by the licensure framework for human pilots.

The proposed framework is based on five pillars: (1) analysis of pilot training and checking regulations with transferable trust anchors; (2) provision of comprehensive and diversified training data to prevent overfitting; (3) introduction of a formal language for machine-interpretable training scenarios; (4) implementation in a simulator-based research setup featuring an electronic flight instructor/examiner; and (5) a validation concept to assess feasibility and effectiveness of the approach.

By embedding human pilot training principles into AI assurance, the approach fosters the trust eventually required for regulatory approval and operational deployment, and highlights the need for broader changes in the aviation system before AI pilots can replace human pilots.