Safety assurance for automated driving systems that can adapt using machine learning: A qualitative interview study.

Introduction: Automated Driving Systems (ADSs) present significant unresolved challenges for traditional safety assurance frameworks. These frameworks did not envisage, nor readily support, automated driving without the active involvement of a human driver, or support safety-critical systems using Machine Learning (ML) to modify their driving functionality during in-service operation.

Method: An in-depth qualitative interview study was conducted as part of a broader research project on safety assurance of ADSs that can adapt using ML. The objective was to capture and analyze feedback from leading global experts, from both regulatory and industry stakeholders, with the key objectives of identifying themes that could assist with the development of a safety assurance framework for ADSs, and providing a sense of the level of support and feasibility for various safety assurance concepts relevant to ADSs.

Results: Ten themes were identified from an analysis of the interview data. Several themes support a whole-of-life safety assurance approach for ADSs, with strong support for ADS developers to be required to produce a Safety Case, and for ADS operators to maintain a Safety Management Plan throughout an ADSs operational life. There was also strong support for in-service ML-enabled changes to be allowed within pre-approved system boundaries, although there were mixed views on whether human oversight of such changes should be required. Across all themes identified, there was support for progressing reform within current regulatory frameworks, without requiring wholesale changes to current frameworks. The feasibility of some themes was identified as presenting challenges, particularly with the ability for regulators to develop and maintain an appropriate level of knowledge, capability and capacity, and with the ability to effectively articulate and pre-approve boundaries within which in-service changes can occur without additional regulatory approval.

Conclusions: Further research on the individual themes and findings would be beneficial to support more informed reform decisions.