The institute’s concept for motion demand planning and generating control inputs for the vehicle motion execution controller uses a two-step approach. In the first step, a smooth path for high driving comfort is planned. The real-time capable OPP is based on an efficiently solvable optimization problem, which minimizes the curvature of splines representing the path within given road boundaries (see figure below). Furthermore, the moving horizon of the OPP enables the optimization to take environmental changes into account. The OPP was experimentally quantified for saving energy while traveling along a given road with an electric vehicle.
In the second step, a velocity profile optimization over the planned path is performed to find a trade-off between time and energy optimality. This trade-off can be directly chosen by the driver. With the help of a dynamic programming framework, the global optimum of the complex optimization problem with nonlinear constraints based on the vehicle’s physical limits is found. The velocity profile optimization is capable of taking environmental changes into account. Subsequently, the planned path together with the velocity profile is provided as input to the PFC. A nonlinear model-based PFC combines a geometric, i.e., time-independent, formulation with a demand supervisor. Based on a geometric vehicle model, the supervisor limits commanded demands, if necessary, to maintain the feasibility of the path following task. The algorithm enables the tracking of predefined paths with high accuracy and assures vehicle stability at all relevant vehicle states. The framework has been extensively tested and validated in real world tests of automated driving with the ROboMObil. Within the institute, a framework for alternative learning-based control approaches has been developed recently. The used method relies on model-free reinforcement learning (RL), which identifies the control law from interaction with the system. The training process uses multiphysical models, e.g. a high-fidelity vehicle model implemented in Modelica (see figure below). To include the system dynamics into the Python-based RL framework, the multiphysical model is incorporated as FMU.