Behaviour cloning of a Model Predictive Controller for Path Tracking applications

This Master's thesis spans a wide range of fields. Starting from the control domain, a comprehensive review of Model Predictive Control (MPC) was conducted. Subsequently, more emphasis was placed on automotive applications, studying the main kinematic and dynamic models of a vehicle as outlined in Rajamani's texts. The focus then shifted to lateral models, forming the foundation for the subsequent MPC development. The initial phase concluded with an extensive literature review on artificial neural networks (ANNs) and their primary applications, aiming to identify the most suitable one for my work. The thesis's objective is to experimentally demonstrate how computational limitations in the automatic control domain can be easily overcome through the use of simple machine learning techniques. Unfortunately, scientific literature on behavior cloning for MPCs is scarce, particularly in its path tracking applications. Since path tracking tasks typically fall under "fast dynamics," model-based approaches tend to be computationally heavy, and hardware may not always be sufficiently powerful. In brief, the ultimate goal was to show that Multilayer Perceptrons (MLPs) could not only replace a complex control architecture but also enable the deployment of computationally unaffordable model-predictive solutions through behavior cloning and offline optimization. The specific scenario addressed in this work is the lane-keeping problem of a scaled sensor-rich autonomous vehicle (SSRAV) by Quanser, provided by the University of Surrey. Previous path-tracking solutions for this control problem included conventional controllers (PI), pole placement (LQR), and MPC. Unfortunately, to get a real-time application, a downsizing was necessary. Prediction horizon got cut by half and longitudinal velocity was cut off from the online data. Nonetheless the MPC still outperformed the previous two architectures in terms of key performance indicators (KPIs). Given the complexity of the task, establishing a coherent criteria for progression was crucial. The first part of the work involved developing a general toolchain capable of quickly and optimally setting up, training, and evaluating MLPs. Since the Quanser environment uses MATLAB as an interface, the generic toolchain also incorporates Python scripts directly called from MATLAB to expedite training. Significant effort was devoted to enabling automated deployment of the MLP created after the Python process on Simulink, with a primary challenge being the management of size-variable signals due to the generalization. The second part of coding involved developing a more complex MATLAB environment allowing users to create a set of expert MPCs for testing and deployment (for different speeds and prediction horizon), collecting data, and ultimately training, evaluating, and deploying the resultings MLP. This specific environment of course incorporates the previously created toolchain. Throughout the coding and simulation process, the simulation bench was reorganized and simplified to streamline the code both on the simulation side and the experimental one. Additionally, the former incremental algorithm estimating the SSRAV position along the path was replaced by one that does not accumulate errors leading us to test back the former-downsized experimental MPC in order to have a consistent comparison. In the end, to underscore the limitations and drawbacks of the MLP method, a literature review oriented towards general imitation learning and its variants was conducted.