Content area
To successfully integrate autonomous vehicles as a mode of transportation, we must test these systems against their end-to-end requirements. When AVs and humans (pedestrians, bicyclists, human-driven vehicles) share the road, they must follow the same traffic rules. Using the common traffic rules (written for human drivers) as an engineering requirement poses a challenge due to the ambiguity of the natural language. On the other hand, an AV is fundamentally different from a human, and a simple road test is not sufficient to assess an AV's compliance and skills. This calls for automated, systematic and scalable testing techniques. The focus of this dissertation is on simulation-based testing of the traffic-rules requirements.
This dissertation develops techniques for systematic exploration of the test-case space of autonomous vehicles based on two crucial concepts: complexity and coverage. Here, these concepts are formalized with respect to the traffic rules requirements. The efficiency of finding bugs is improved by incrementally increasing the complexity of a test-case, namely making it harder for an AV to pass a test-case. On the other hand, the diversity of a test-suite is improved by guiding the test-case generation towards increasing the coverage of a test-suite. The framework for formalization of the traffic rules is made more amenable to vetting by the authorities and regulators by narrowing the gap between human intuition and machine language. This is achieved by formalizing traffic rules in first-order logic (FOL) which allows modeling objects and predicates.