Trajectory Planning for Autonomous Vehicles combining Nonlinear Optimal Control and Supervised Learning

Abstract This paper considers computationally efficient planning of reference trajectories for autonomous on-road vehicles in a cooperative setting. The basic element of the approach is the notion of so-called maneuvers, which allow to cast the nonlinear and non-convex planning task into a highly structured optimal control problem. This can be solved quite efficiently, but not fast enough for online operation when considering nonlinear vehicle models. Therefore, the approach proposed in this paper aims at approximating solutions using a supervised learning approach: First, training data are generated by solving optimal control problems and are then used to train a neural network. As is demonstrated for a cooperative overtaking maneuver, this approach shows good performance, while (contrasting approaches like reinforcement learning) requiring only low training effort.