Preliminary Evaluation of Path-aware Crossover Operators for Search-Based Test Data Generation for Autonomous Driving

Abstract

As autonomous driving gains attraction, testing of autonomous vehicles has become an important issue. However, testing in the real world is not only dangerous but also expensive. Consequently, a virtual test method has emerged as an alternative. Recently, a novel testing technique based on Procedural Content Generation (PCG) and Genetic Algorithm (GA), As-Fault, has been proposed to test the lane-keeping functionality of autonomous vehicles. This paper proposes new crossover operators for AsFault that can better preserve the coupling between genotype (representations of road segments) and phenotype (occurrences of interesting self-driving behaviour). We explain our design intentions and present a preliminary evaluation of the proposed operators using the Simulink autonomous driving simulator. We report promising early results: the proposed operators can lead not only to Out of Bound Episodes (OBEs) but also causes more vision errors in the simulation when compared to the original.