A virtual driving instructor that assesses driving performance on par with human experts

Abstract

The advent of virtual driving instructors has the potential to revolutionize driver education by providing real-time, unbiased feedback to learner drivers. This paradigm shift aims to mitigate the innate subjectivity associated with human evaluations. Our research focused on the creation of a virtual driving instructor capable of assessing a learner driver’s performance in real-time, with an emphasis on eliminating the inherent biases associated with human evaluations. Our approach involved the development of a rule-based assessment system, employing a multi-agent system based on the subsumption architecture. Each agent in the system was tasked with assessing a specific aspect of driving performance. Additionally, we utilized a knowledge graph to maintain a continuous understanding of the situational context, further enhancing the system’s assessment capabilities. We posited that our system, given its methodical structure and objective rule-based framework, would be able to accurately and objectively assess various driving scenarios. Further, we hypothesized that our system’s performance would be on par with expert human evaluations. The validation of our system was conducted using real driving sessions in simulators with actual students. The system was tested on various scenarios including intersections, roundabouts, and overtakes. The assessment results aligned closely with expert consensus, showcasing the system’s capacity to match the evaluative precision of human experts.