Author:

Abstract:

The demand of industry to increase the autonomy of mechatronic systems drives the research and development of digital technologies to enable performant autonomous motion planning and control in a large range of tasks. Among the techniques which have emerged to tackle mechatronic motion planning and control tasks, model predictive control (MPC) is a popular model-based method which offers the benefits of safe, robust, and interpretable motion trajectory generation and tracking. MPC is capable of explicitly accounting for actuator limits and other constraints through the use of numerical constrained optimization. Although MPC has proven successful in numerous applications, its use is most often restricted to systems with slow or linear dynamics, to applications with short prediction horizons and simple objectives, or to reference tracking problems with an (almost) feasible reference. Especially considering the limited available computation resources on compact autonomous vehicles, the computationally costly online evaluation of MPC is still a hurdle for the adoption of the method in more complex motion planning tasks. Moreover, the tuning of MPC is most often a matter of engineering experience, and a cumbersome procedure of trial-and-error. This thesis aims at making the MPC methodology more suitable for the online motion planning tasks of three types of autonomous vehicles: agile drone flight, truck-trailer autonomous mobile robot maneuvering, and highway lane changing of a self-driving car. To this end, this thesis addresses five aspects of the MPC methodology: (1) the modeling of vehicle dynamics for use in an optimal control problem formulation, (2) the modeling of the environment in which the vehicle moves, (3) the strategy of online replanning with a receding horizon, (4) the automatic real-time selection of different constraint sets to reflect discrete decisions made by the control system, and (5) the choice of the tuning parameters in the control policy. Firstly, for the modeling of vehicle dynamics, this thesis explores the modeling and parameter identification of drone dynamics. An approximate model which allows a safe and simple procedure to identify the parameters solely based on flight data is proposed, evaluated, and critically assessed in comparison with standard models from literature. Secondly, the modeling of the environment is covered for all three applications, by applying a convex free space formulation to represent the available space between obstacles and other forbidden regions. Simulation and real-world experiments demonstrate how this approach translates complex motion planning problems to simpler subproblems which are efficiently and robustly solved online. For the car driving on a highway, a tunable risk taking factor is introduced in the free space formulation, to enable the learning of a balance between assertiveness and courteousness of the autonomous vehicle. Thirdly, this thesis presents an asynchronous update scheme for motion planning with MPC, ASAP-MPC, which handles varying computational delays while still solving the motion planning problem to full convergence on an as-soon-as-possible basis. The implementation of this approach on compact embedded hardware proves its effectiveness. A comparison with state-of-the-art MPC update schemes shows that ASAP-MPC is capable of handling more complex motion planning scenarios. Fourthly, a control software structure is introduced, consisting of a two-layered discrete and continuous control policy of which the discrete and continuous components interact through the states of a finite state machine. The structure is shown to make an autonomous vehicle successfully perform lane changes in close interaction with an approaching vehicle, and to recover safely when the intended discrete decision is infeasible. Finally, the control software structure is integrated in the OpenAI Gym toolkit, to prepare it for use in reinforcement learning and imitation learning training experiments, such that, in future work, the driving behavior of the two-layered control policy can be learned from expert demonstrations.