Objekterkennung von Fahrbahnbegrenzungen für Reinforcement Learning basierte Pfadfolgeregelung

Abstract

The perception of the environment is a crucial task in autonomous driving and needs to be solved in real-time in order to enable a safe and efficient operation. Especially for the motion planning application of an autonomous vehicle, a precise model of the environment needs to be provided. In this thesis, a neural network-based path detection method is applied for the ROboMObil research vehicle of the German Aerospace Center. This setup enables the camera-based perception of path boundaries and sets the foundation for the real-time application of the ROboMObil’s path planning module. The experimental perception results on a German highway demonstrate that the path boundaries are detected precisely. Further, the motion of the vehicle needs to be controlled in a robust way once a path has been provided by the path planning module. In this thesis, an existing framework for the reinforcement learning based path following control (PFC) of the ROboMObil is extended by dynamics randomization, i.e. sampling different values of selected model parameters during the training, which enables the learning of robust agents. The simulation results show that PFC agents trained with randomized dynamics obtain a higher robustness against varying environmental conditions and model uncertainties.