Early Concept Evaluation of a Runtime Monitoring Approach for Safe Automated Driving
Being used in key features, such as sensing and intelligent path planning, Artificial Intelligence (AI) has become an inevitable part of automated vehicles (AVs). However, their usage in the automotive industry always comes with a “label” that questions their impact on the overall AV safety. This paper focuses on the safe deployment of AI-based AVs. Among the various ways for ensuring the safety of AI-based AVs is to monitor the safe execution of the system responsible for automated driving (i.e., Automated Driving System (ADS)) at runtime (i.e., runtime monitoring). Most of the research done in the past years focused on verifying whether the path or trajectory generated by the ADS does not immediately collide with objects on the road. However, as we will show in this paper, there are other unsafe situations that do not immediately result in a collision but the monitor should check for them. To build our case, we have looked into the National Highway Traffic Safety Administration (NHTSA) database of 5.9 million police-reported light-vehicle accidents and categorized these accidents into five main categories of unsafe vehicle operations. Furthermore, we have performed a high-level evaluation of the runtime monitoring approach proposed in [1], by estimating what percentage of the total population of 5.9 million of unsafe operations the approach would be able to detect. Lastly, we have performed the same evaluation on other existing runtime monitoring approaches to make a basic comparison of their diagnostic capabilities.