Augmenting anomaly detection for autonomous vehicles with symbolic rules

Abstract

My research investigates the issues in anomaly detection as applied to autonomous driving created by the incompleteness of training data. I address these issues through the use of a commonsense knowledge base, a predefined set of rules regarding driving behavior, and a means of updating the base set of rules as anomalies are detected. In order to explore this problem I have built a hardware platform that was used to evaluate existing anomaly detection developed within the lab and that will serve as an evaluation platform for future work in this area. The platform is based on the open-source MIT RACECAR project that integrates the most basic aspect of an driving autonomous vehicle - lidar, camera, accelerometer, and computer - onto the frame of an RC car. We created a set of rules regarding traffic light color transitions to test the car's ability to navigate cones (which represent traffic light colors) and detect anomalies in the traffic light transition order. Anomalies regularly occurred in the car's driving environment and its driving rules were updated as a consequence of the logged anomalies. The car was able to successfully navigate the course and the rules (plausible traffic light color transitions) were updated when repeated anomalies were seen.