Concrete scenario generation with a focus on edge cases for the safety assessment of highly automated vehicles

Abstract

Modern autonomous vehicles will undoubtedly include machine learning and probabilistic techniques. These algorithms used within a non-deterministic world significantly complicate the safety assessment process. In addition, the HAVs must handle the responsibility of safe navigation as the driver is optional, so there will be no one to guarantee safety. The first contribution of this thesis is addressing the challenges of testing HAVs, and the thesis presents a comprehensive set of issues regarding this topic.

Recently, academia and industry have suggested that scenario-based testing (SBT) could be complementary to large-scale on-road testing as its focus is biased toward relevant high-risk scenarios. However, a naive approach encounters issues such as parameter space explosion, potentially requiring an infeasible number of scenarios to find challenging situations. As a second contribution, the thesis presents two novel scenario generation methods that enable the evaluation of SUT by biasing the learning towards the edge cases. The methods utilise the parameter space based on expert knowledge for generating scenarios in a pedestrian crossing traffic context.

The expert knowledge-based parameter space may not represent the realistic behaviour of the traffic participants. The thesis introduces an edge-case focused scenario generation approach based on a parameter space built from real-world data as a third contribution. We use a novel method to convert collected raw data into parametric representations of relevant scenarios. Finally, we generate realistic concrete scenarios, combining our data-driven parameter space with a reinforcement learning (RL) based method. Overall, the thesis introduced a method that can generate risky scenarios as well as ensures that the generated scenarios are closer to the real-world data.