Model Learning and Tactical Maneuver Planning for Automatic Driving

Abstract

Tactical maneuver planning is one of the key enablers for automated driving. The challenges include complex situations in urban areas and the uncertain behavior of other road users. In this paper, we present an approach to model the decision problem of tactical maneuver planning as a Markov decision process (MDP) for a two-lane road. It is shown how this model can be used to make tactical maneuver decisions on a three-lane road without increasing the complexity of the MDP. Furthermore, it is shown how the model can be learned and improved in a three-lane simulation environment using real-world experience. The results show that the learned model represents the environment better than the manually modeled MDP and that a significantly better driving strategy is calculated based on this.