JARVIS D 5.3 Whitepaper on AI Assurance

This document primarily aims to collect the lessons learned on AI assurance identified during the development of artificial intelligence algorithms within the JARVIS project.
The project aims to deliver seven different Digital Assistant (DA) solutions based on AI, one for on-board aircraft operations, one supporting four different tasks of air traffic controllers, and four supporting airport operators.

The DAs and their feature are based on different techniques, including supervised and unsupervised learning, as well as Reinforcement Learning and Automated Reasoning–based approaches. The lessons learned focus on some of the aspects that JARVIS AI developers found most challenging at the individual feature level, namely: the definition of the Operational Design Domain (ODD), Data management aspects, and the properties of robustness, generalization, and explainability. In addition, the document provides some considerations on safety properties at the system level.

The lessons learned further focus on two techniques not currently addressed in the EASA Concept Paper, namely Automated Reasoning–based approaches and Reinforcement Learning, which were also discussed in depth with EASA during the development activities. The various lessons learned were collected through workshops involving developers, held in several sessions over the two and a half years of the project duration. Whenever possible, the lessons learned are accompanied by examples to clarify their intended meaning.

The JARVIS project, a SESAR Joint Undertaking initiative (Grant 101114692), ran from 2023 to 2026 and developd three digital assistants for airborne, air traffic control and airport operations. Within workpackage 5 (WP 5) lessons learned on foundational AI challenges dataset creation, AI design assurance and Human-AI teaming were collected.