Testing autonomous driving systems against sensor and actuator error combinations

Pascal Minnerup, Alois C. Knoll

Abstract: Simulation as a tool for validating the behavior of an autonomous driving system can include sensor and actuator errors. A combination of such errors might lead to undesired behavior not occurring with a single error. In this paper we present a new concept for efficiently covering these error combinations. We propose to divide the simulation sequence into small separately simulated pieces that can be assembled to cover more patterns of error combinations. This way, parts of a simulation sequence that are similar for multiple error patterns are run only once. Our experimental results show that the concept can find weaknesses of planning and control systems and is faster than running a separate simulation sequence for each error pattern.