Direct Data-Driven and Model-Based Control Design for an Autonomous Bicycle

Autonomous bicycles constitute challenging benchmark systems for control, due to their nonlinear, non-holonomic, and in general, underactuated, open-loop unstable dynamics. Traditional model-based controllers such as proportional-integral-derivative (PID) controllers and linear quadratic regulators (LQRs) can stabilize the bicycle, but rely on simplified models that may not capture unmodelled and time-varying effects. In contrast, recent direct data-driven control methods based on Willems’ fundamental lemma bypass explicit modelling, yet typically assume linear time-invariant dynamic systems and require persistently exciting inputs that are difficult to apply safely on unstable systems. This thesis investigates how traditional model-based and direct data-driven control methods can be used, and combined, to balance and guide an autonomous bicycle using mainly steering actuation as input. First, PID, LQR, fuzzy controller, feedback linearization (FL), and direct data-driven controllers are designed and compared in high-fidelity simulations and experiments on an autonomous bicycle. The results show that classical model-based controllers provide strong baselines, while direct data-driven controllers can enhance performance when combined with classical controllers. Second, a unified framework is proposed in which an inner-loop FL controller stabilizes and partially linearizes the bicycle, and an outer-loop direct data-driven controller operates on the FL-stabilized system. Two different types of direct data-driven methods are evaluated in this setting: a static, nonlinear controller and the Data-enabled Policy Optimization (DeePO) algorithm. Third, the DeePO algorithm is analysed and modified to mitigate state perturbations, leading to a perturbation-free variant studied on LTI systems. Finally, a model-based PID–MPC trajectory tracking scheme is compared with a data-driven framework relying on Data-enabled Predictive Control (DeePC) for trajectory tracking, combined in a cascade architecture with the FL-DeePO setup. Simulations show that while PID–MPC achieves better tracking accuracy, the data-driven cascade attains successful trajectory tracking without relying on an explicit dynamic model.