This thesis explores control design methodologies within the direct data-driven learning paradigm, with a particular focus on iterative strategies that improve the control law directly from experimental data, avoiding explicit model identification. The proposed approaches are validated in robotic applications involving mobile robots and manipulators, with the aim of achieving good closed-loop performance and high computational efficiency. A first contribution introduces a kernel-based method for learning feedforward controllers within an Iterative Learning Control (ILC) framework for nonlinear systems. Unlike traditional ILC schemes, the control update is purely data-driven. The method is compared with a conventional ILC approach and a neural-network-based controller, showing improved tracking performance and convergence in simulation on a path-following problem for a unicycle robot. The following chapters present imitation learning applications, where a high-performance expert controller, specifically a nonlinear Model Predictive Controller, is used to generate training data. These data are employed to train a kernel-based Linear Parameter Varying controller and a controller based on Spiking Neural Networks (SNNs), which are evaluated in simulation, being applied to control the pose of the end-effector of a manipulator. Finally, the thesis describes a real-world implementation of an SNN-based controller for wheel speed control of a mobile robot, demonstrating transferability to real hardware. Overall, the results show that the proposed methods represent valid data-driven alternatives to model-based control strategies, enabling comparable performance and higher control frequencies. Future perspectives include closed-loop stability analysis, formal robustness evaluation, and extensions toward adaptive online controllers.
In questa tesi si esplorano metodologie di progettazione del controllo nel paradigma del direct data-driven learning, con particolare attenzione a strategie iterative che migliorano la legge di controllo direttamente dai dati sperimentali, evitando l’identificazione esplicita del modello. Gli approcci proposti sono validati in applicazioni robotiche su robot mobili e manipolatori, con l’obiettivo di ottenere buone prestazioni in anello chiuso e un’elevata efficienza computazionale. Un primo contributo introduce un metodo basato su kernel per l’apprendimento di controllori feedforward in un framework di Iterative Learning Control (ILC) per sistemi non lineari. A differenza degli schemi ILC tradizionali, l’aggiornamento del controllo è puramente data-driven. Il metodo è confrontato con un ILC convenzionale e con un controllore ILC basato su rete neurale, mostrando in simulazione migliori prestazioni di tracking e convergenza in un problema di path-following di un robot uniciclo. I capitoli successivi presentano applicazioni di imitation learning, in cui un controllore esperto ad alte prestazioni, nello specifico un Model Predictive Controller non lineare, è utilizzato per generare dati di addestramento. Tali dati sono impiegati per addestrare un controllore Linear Parameter-Varying basato su kernel ed un controllore basato su Spiking Neural Networks (SNN), valutati in simulazione applicati al controllo di un manipolatore. Infine, la tesi descrive un’implementazione reale di un controllore SNN per il controllo della velocità delle ruote di un robot mobile, dimostrandone la trasferibilità su hardware reale. Nel complesso, i risultati mostrano che i metodi proposti rappresentano valide alternative data-driven ai controlli basati su modello, consentendo prestazioni comparabili e frequenze di controllo più elevate. Prospettive future includono analisi di stabilità in anello chiuso, valutazioni di robustezza formale ed estensioni verso controllori adattivi online.
Learning-Based Data-Driven Control with Applications in Robotics: from Kernel Methods to Neuromorphic Approaches / Vella, Francesco. - (2026 Mar).
Learning-Based Data-Driven Control with Applications in Robotics: from Kernel Methods to Neuromorphic Approaches
VELLA, FRANCESCO
2026-03-01
Abstract
This thesis explores control design methodologies within the direct data-driven learning paradigm, with a particular focus on iterative strategies that improve the control law directly from experimental data, avoiding explicit model identification. The proposed approaches are validated in robotic applications involving mobile robots and manipulators, with the aim of achieving good closed-loop performance and high computational efficiency. A first contribution introduces a kernel-based method for learning feedforward controllers within an Iterative Learning Control (ILC) framework for nonlinear systems. Unlike traditional ILC schemes, the control update is purely data-driven. The method is compared with a conventional ILC approach and a neural-network-based controller, showing improved tracking performance and convergence in simulation on a path-following problem for a unicycle robot. The following chapters present imitation learning applications, where a high-performance expert controller, specifically a nonlinear Model Predictive Controller, is used to generate training data. These data are employed to train a kernel-based Linear Parameter Varying controller and a controller based on Spiking Neural Networks (SNNs), which are evaluated in simulation, being applied to control the pose of the end-effector of a manipulator. Finally, the thesis describes a real-world implementation of an SNN-based controller for wheel speed control of a mobile robot, demonstrating transferability to real hardware. Overall, the results show that the proposed methods represent valid data-driven alternatives to model-based control strategies, enabling comparable performance and higher control frequencies. Future perspectives include closed-loop stability analysis, formal robustness evaluation, and extensions toward adaptive online controllers. I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


