LAUSR

Abstract Safe autonomous passing of intersections with mixed traffic, including human drivers and autonomous vehicles, is challenging. We propose a tailored approach that provides guarantees despite uncertainties fusing learned models and model predictive control. A single autonomous vehicle is controlled by the predictive controller via acceleration and steering angle without assumption of a global controller. Each maneuver of the human behaviour is modeled with a neural network, which enters the predictive controller formulation as a constraint. As an example, we consider a single autonomous vehicle on an unsignalized intersection, which gives right-of-way to a human-driven vehicle. We show how human driving behavior can be modeled based on real recorded trajectory data and implemented in the proposed predictive control approach by dynamically changing the constraints of the optimization problem.