Learning autonomous vehicle passengers’ preferred driving styles using g-g plots and haptic feedback

In this work, we present and experimentally validate a framework for learning an autonomous vehicle passenger’s preferred driving style. The study is performed with N = 3 human subjects in a vehicle simulator that consists of a 3-DOF motion simulator, providing feelings of longitudinal and lateral acceleration to the passenger, an automatically controlled steering wheel, providing information about the steering controller behavior, and a computer monitor, providing a virtual rendering of the view through the windshield. The vehicle controller is designed to track speeds while satisfying limits on the maximum allowable longitudinal and lateral accelerations. These accelerations are related to the passenger’s preferences and are represented as a surface on a g-g plot. The passenger’s preferences are learned from comfort labels provided by the passenger, which correspond to positive and negative assessments of the vehicle’s current driving behavior. In the framework that we present, these labels directly change the corresponding parametrization of the g-g plot, thereby modifying the limiting constraints to be enforced by the controller on-line, which leads to a change in the behavior of the vehicle. The collected data supports the hypothesis that there is a personalized driving style preference, and also shows that our proposed preference-learning scheme converges to a preferred driving style.