Testing Automated Vehicles Against Actuator Inaccuracies in a Large State Space

Inaccurate actuator responses affect the behavior of an autonomous driving system. A disadvantageous combination of such inaccuracies might lead to a collision, but is hard to test in advance due to the exponentially large number of possible combinations. This paper introduces STARVEC, a tool to test autonomous driving systems for undesired behaviors in the presence of sensor and actuator inaccuracies in a simulation environment. It stores intermediate states of the simulation and uses these states to efficiently explore the space of possible behaviors. Each step continues with the execution of the state with the highest distance to its neighbors. Thus, the potentially large space of reachable states is covered fast and increasingly dense. The approach is applied to an autonomous parking system with inaccurate actuators and its performance is compared to a Monte-Carlo algorithm and a previous prototype.