My final DPhil thesis is now available for download

Abstract: Intelligent vehicles with automated driving functionalities provide many benefits, but also instigate serious concerns around human safety and trust. While the automotive industry has devoted enormous resources to realising vehicle autonomy, there exist uncertainties as to whether the technology would be widely adopted by society. Autonomous vehicles (AVs) are complex systems, and in challenging driving scenarios, they are likely to make decisions that could be confusing to end-users. As a way to bridge the gap between this technology and end-users, the provision of explanations is generally being put forward. While explanations are considered to be helpful, this thesis argues that explanations must also be intelligible (as obligated by the GDPR Article 12) to the intended stakeholders, and should make causal attributions in order to foster confidence and trust in end-users. Moreover, the methods for generating these explanations should be transparent for easy audit. To substantiate this argument, the thesis proceeds in four steps: First, we adopted a mixed method approach (in a user study N = 101) to elicit passengers’ requirements for effective explainability in diverse autonomous driving scenarios. Second, we explored different representations, data structures and driving data annotation schemes to facilitate intelligible explanation generation and general explainability research in autonomous driving. Third, we developed transparent algorithms for posthoc explanation generation. These algorithms were tested within a collision risk assessment case study and an AV navigation case study, using the Lyft Level5 dataset and our new SAX dataset—a dataset that we have introduced for AV explainability research. Fourth, we deployed these algorithms in an immersive physical simulation environment and assessed (in a lab study N = 39) the impact of the generated explanations on passengers’ perceived safety while varying the prediction accuracy of an AV’s perception system and the specificity of the explanations. The thesis concludes by providing recommendations needed for the realisation of more effective explainable autonomous driving, and provides a future research agenda.