Towards Modelling Sensor Failures in Automotive Driving Simulators

Testing autonomous vehicles is a costly and tedious task, but essential to ensure their safe operation. In order to cover diverse scenarios and stage emergency situations like accidents safely, many manufacturers rely on automotive driving simulators. These simulators are virtually mirroring real driving situations by simulating an ego vehicle and its sensors to generate perceptual output for training or testing purposes of the autonomous driving software. Currently, a "functional paradigm" is followed by creating a realistic representation of the environment and running an ideal simulation. However, in reality sensors do not generate ideal outputs, as they are victim to various environmental effects like interfering signals, reflections, blockage and naturally ageing and wear. Furthermore, the environment can be flawed, e.g., traffic signs may be damaged or obscured. Consequently, we suspect that in the future a shift to a "failure paradigm" will become necessary where simulators include faults and failures of the infrastructure, the ego vehicle and its components. With that, a more realistic representation of driving in the actual world could be achieved, producing valuable data for training and testing of fault detection mechanisms.

With this paper we aim to make first steps towards the direction of a failure paradigm by the example of automotive radar. We study existing literature regarding the degradation of the radar function and propose four generalized failure models. We present the partial implementation of these in the open-source simulator CARLA and discuss experimental results as well as future contributions to this topic.

Keywords: Autonomous driving, Automotive, Driving simulators, Safety, Security, Reliability, Resilience, Failure behavior, Testing.

Download PDF