Improving Driving Performance on a Scaled Autonomous Vehicle with Model Predictive Control

This master thesis discusses the modeling and controller design of a 1/10 scaled ground vehicle. In particular, we explore how nonlinear predictive control influences performance in our automotive application. This control methodology is particularly effective when dealing with multiple-input multiple-output systems, as is the case in our thesis, and it allows systematic incorporation of physical and operational constraints in the control design. The vehicle model must balance accuracy and computational efficiency to ensure real-time feasibility, as the chosen model complexity directly influences different performance and goals that can be achieved such as trajectory tracking precision or lap time optimization. Different models are explored, starting from the simplest kinematic bicycle model to the double dynamic bicycle model. In the case of the dynamic model, the benefits of modeling longitudinal and lateral tire forces are considered in different ways. This makes space for better performance in the high speed scenarios and curvature track, but is not suitable for low-speed scenarios. Initially, the models in the time domain are studied, but later on focus is shifted to the spatial domain formulation. This allows easier description of the constraints, and to provide a velocity reference not dependent on time, but rather on the position on the track. The model and the controller were evaluated on a predefined track. This configuration allows the computation of reference trajectories offline, providing a reliable benchmark for trajectory tracking and control performance assessment. All simulations are conducted in Python utilizing the Acados framework, which offers efficient tools for implementing real time Model Predictive Controllers. The tracks used in the simulation are elliptical and S-shaped configurations, selected for their ability to highlight different aspects of the controller’s performance, such as cornering stability, trajectory tracking accuracy, and response to curvature variations. Later on, the control architecture is integrated into a ROS2-based system and validated in the Gazebo simulator on tracks from Bosch Future Mobility Challenge (BFMC) environment. The BFMC is an international competition organized by Bosch Engineering Center Cluj where teams of university students develop and implement autonomous driving and connectivity algorithms on 1/10 scale vehicles. Incorporating tracks inspired by this competition allows the simulation framework to better reflect real-world testing conditions in urban-like environments. A hardware-in-the-loop (HIL) configuration is additionally used, where a Raspberry Pi 5 runs the MPC controller. This configuration allows assessment of computational feasibility, actuator-rate limitations, communication delays, and integration effects that cannot be captured in purely offline simulations. Thus, the architecture loop bridges the gap between algorithmic design and physical implementation, ensuring that the controller’s performance is validated under realistic constraints imposed by embedded hardware, middleware communication, and physics-based simulation.