Applications of Robust Statistics in Autonomous Driving

Autonomous vehicles have to interact with their environment with the goal to fulfill their tasks while respecting all desired constraints such as not causing dangerous situations, driving comfortable maneuvers, enabling a smooth traffic flow, or avoiding overly polluting driving behavior. All steps require a suitable perception of the environment conditions, such as the estimation of the own position, a prediction of the trajectories of other traffic participants, or the assessment of parameters corresponding to vehicle dynamics. However, classical estimation algorithms are known to be easily distorted by outliers in the data. In addition, apart from rule-based systems, it becomes more convenient to train autonomous agents by machine learning algorithms. Again, such algorithms need to be robust in order to cope with model misspecification or outliers in the data. Robust Statistics is a discipline of statistics which exactly addresses these challenges. This paper provides an extensive and systematic overview of current applications of Robust Statistics in autonomous driving in a unified notation, discusses different notions of the term “robustness” and identifies directions for future work.