Geïntegreerde Leer- en Optimalisatietechnieken voor de Besturing van Autonome Voertuigen

Autonomous driving (AD) technology has seen rapid growth in the automotive industry, driven by demands for enhanced safety, comfort, stability and environmental sustainability. Among the several components that go into an autonomous vehicle, is the controller or the decision making brain. While Nonlinear Model Predictive Control (NMPC) has emerged as a promising advanced control strategy in robotics and autonomous driving by achieving high performance with safety guarantees, four fundamental challenges impede its transfer from simulation to the real world, hindering its industrial adoption: 1) critical real-time implementation due to heavy computational requirements of optimization routines, 2) ease of prototyping on different target platforms, 3) catastrophic performance degradation between simulation and reality, and 4) limited explainability of the NMPC's decision making process and performance in complex closed-loop systems. Recent (blackbox) machine learning advances in AD control demonstrate advantages in complex scenarios and performance recovery. However, they lack rigorous guarantees for explainability, safety, and stability. This thesis bridges this gap by combining optimal model-based control and data-driven approaches for embedded autonomous driving applications, addressing performance optimization, as well as data and computational efficiency. This researches the topic of transferring an AD control module from simulation to the real world (Sim2Real), for efficiently transferring knowledge developed in simulation to the real systems, reducing real experiments time and cost, data collection and manual tuning efforts. The focus is on four main aspects: First, this thesis tackles certifiable, safe, and real-time capable optimal controllers. It introduce RESAFE/COL, a novel transcription approach allowing to efficiently solve continuous-time optimal control problems. It overcomes NMPC safety constraints' scalability issues, in terms of short horizon and no continuous-time constraint satisfaction, while considerably reducing the computational burden and maintaining a high solution accuracy. Second, this thesis presents a fast AD controller verification and validation framework across simulation, hardware, and physical testing using two developed tools AutoNMPC exporter that generates standalone NMPC modules for rapid prototyping on various target platforms, and CosimApp a tool integrating the controller with high-fidelity simulation tools for the vehicle dynamics and environment, enabling seamless integration with real vehicles and sensors through code unification and a modular framework. Third, this thesis proposes employing Digital Twin technologies for robust AD control adaptation to real-world uncertainties through a re-imagined bi-directional V-cycle development. The thesis proposes a Sim With Real flavor to the Sim2Real transfer of parametric controllers. By leveraging simultaneously 1) executable-digital-twins-in-the-loop for parameter exploration and 2) real-world data performance feedback, it establishes an efficient controller autocalibration pipeline significantly reducing experimental time and manual tuning efforts. The D2C2-AUKS method is introduced for blackbox optimization and rapid controller tuning, achieving an autocalibration of the NMPC cost function in under 10 minutes in real-world conditions, and recovering the lost performance. A virtual driver model is also developed for autonomous racing applications, allowing to extract high performance from controller-plant combinations, safely pushing the NMPC to its limits with racing at speeds over 330kph. Fourth, this thesis aims at augmenting classical control development with Explainable AI technologies to elucidate the optimization process and make it clearer for both control engineers and non-experts. An explainable NMPC framework is presented, ExAMPC, providing transparency and interpretability in autonomous decision-making processes and insights into the relationships between system parameters and closed-loop performance metrics. ExAMPC also allows approximating the optimized NMPC's forecast trajectory by maintaining physical insights through an efficient embedding of the time-series. The framework enables efficient prediction of multi-step sequences in the continuous-domain instead of the discrete one, with a reduction in the output dimensionality while still maintaining physical insights. The effectiveness of the proposed contributions is demonstrated through various AD applications on the SimRod vehicle, including real-time NMPC and AI driver-model implementations such as Deep Reinforcement Learning, both in simulation and real-world environments. The optimal controllers are successfully transferred to industrial-grade ECUs and real-time platforms, providing a comprehensive solution for transferring advanced control strategies to real-world autonomous vehicles while ensuring safety, interpretability, performance and computational guarantees.