USING OPERATIONAL DESIGN DOMAIN FOR SAFE AI IN URBAN AIR MOBILITY

Abstract

In engineering of autonomous systems, Operational Design Domain (ODD) is an artifact that specifies the envelop in which a system safely operates. By defining requirements regarding the safe operating conditions, the ODD provides the basis for the verification and validation of the system’s behavior. ASAM contributed to discussion by defining an ODD framework based on the recent industrystandards such as the PAS 1883:2020 ODD taxonomy for an automated driving system (ADS). The framework meant to guide development of ODDs for different applications. Urban Air Mobility (UAM) is a new segment in aviation that is characterized by intelligent and highly interconnected aircraft which will adopt safe artificial intelligence (AI) components. Hence, ODD is one of the central concepts for the upcoming AI-based safety critical systems of UAM. To introduce AI into aviation the European Union Aviation Safety Agency (EASA) issued number of documents including the Concepts of Design Assurance for Neural Networks reports. They suggest using synthesized data for training, validation and testing machine learning (ML) models for AI-based systems. This approach requires predefined scenarios which generate synthetic data. Operational scenarios are the major components of the Concept of Operations (ConOps). Traditionally, ConOps has been used to describe the characteristics of a proposed system from the user’s perspective. However, using ConOps alone in developing an AI-based system seems insufficient due to the complex interdependencies This paper describes a systematic approach for combining scenarios from ConOps to capture specific operational ranges and limitations to define an ODDs. It further links the ConOps, scenarios, ODD to synthetic data generation for training, validation and testing of ML models. Based on an UAM-use-case, the corresponding ConOps scenarios are defined and are used as a basis to derive the ODD. This implies to transfer the application of ODD from the automotive field to the UAM. This work is a baseline to develop adaptive ODD in order to implement safe AI for robust and failure tolerant systems.