Pushing Cars' Limits: Exploring Autonomous Technologies in the Formula SAE Driverless Competition

This thesis details the design and implementation of a complete software stack for a first-generation Formula SAE Driverless electric vehicle, establishing a foundational architecture for navigating unknown, cone-delineated tracks. The system's architecture is built upon a foundation of classical robotics algorithms while exploring the integration of modern machine learning techniques. The perception pipeline utilizes a stereocamera with a custom-trained YOLOv11n object detector and an ORB-enhanced stereo matching algorithm to reconstruct the 3D track environment. For control, a dynamic vehicle model was formulated to serve as the predictive core for a Non-linear Model Predictive Control (NMPC) framework designed with the acados toolkit. A key contribution of this work is the exploration of a hybrid control strategy using Reinforcement Learning. A Proximal Policy Optimization (PPO) agent was trained to perform online hyperparameter tuning of a PD path-following controller, combining machine learning with a classical, interpretable framework. Experimental validation shows the potential efficacy of the training setup and demonstrates that the PPO agent successfully learns its optimization task, while also highlighting the stability limits of the underlying classical controller. This work establishes a complete software foundation for autonomous racing and provides clear, data-driven insights for future development in high-performance perception and control.