Over the last decades, autonomous vehicle technologies have been advancing steadily. The versatility, the technological success and the cost reduction led the spread of drones and other self-driving vehicles in many research fields. Among the main functionalities that autonomous vehicles need are environmental perception, path planning and especially trajectory tracking. The capability to effectively follow a reference trajectory is critical to the successful completion of an autonomous vehicle’s mission. Nonetheless, the lack of accurate dynamic models prevents the design of control systems ensuring good tracking performance. Specifically, Unmanned Aerial Vehicles (UAVs), commonly known as drones, have nonlinear, dynamically coupled, and difficult to establish models. In recent times, researchers have resorted to machine learning algorithms to identify dynamical systems with limited prior knowledge on the model structure. Above all learning tools, Bayesian nonparametric modeling has shown strong potential for its flexibility and its efficiency in integrating available information on the model. In specific, Gaussian processes are Bayesian nonparametric approaches that have gained huge popularity because they can provide uncertainty estimates for their predictions. The goal of this thesis is to design a control system targeted for a specific class of nonlinear systems called differentially flat systems (including quadrotor UAVs). The control system aims at achieving satisfactory tracking performance by exploiting a Gaussian process as a learning tool. As a novelty compared to the literature on the subject, this work explicitly considers the inaccuracy of the Gaussian process inputs and evaluates how it damages identification and trajectory tracking. In addition, the thesis includes a novel approach to address an open problem in control theory: the robustness of learning-based systems. In conclusion, the control system is tested in simulation on the lateral dynamics of a quadcopter. The obtained results demonstrate the effectiveness of the control system in achieving good tracking despite the limitations brought by the input uncertainty.
Negli ultimi decenni, le tecnologie dei droni e dei veicoli autonomi hanno subito uno sviluppo esponenziale. Il successo tecnologico, la versatilità applicativa, e la riduzione dei costi hanno portato alla diffusione di tali tecnologie in molti campi di ricerca. I veicoli autonomi si basano su tre funzionalità cardine: percezione dell’ambiente circondario, pianificazione del percorso e tracciamento di traiettoria. La capacità di seguire efficacemente una traiettoria di riferimento è fondamentale per completare con successo la missione di un veicolo autonomo. Tuttavia, la mancanza di modelli dinamici accurati impedisce la progettazione di sistemi di controllo che garantiscano buone prestazioni di tracciamento. In particolare, gli Unmanned Aerial Vehicles (UAV), noti anche come droni, sono sistemi non lineari, accoppiati dinamicamente e difficili da indentificare. Recentemente, i ricercatori hanno ricorso ad algoritmi di machine learning per identificare i sistemi dinamici di cui ci sia solo una conoscenza preliminare limitata. Tra le diverse tecniche di learning, i metodi fondati sulla teoria di Bayes hanno mostrato forte potenziale nell’identificare efficacemente i modelli dinamici con un approccio probabilistico. In particolare, i processi Gaussiani sono metodi bayesiani non parametrici che hanno ottenuto una grande popolarità in quanto possono generare stime di incertezza riferite alle loro predizioni. L’obiettivo di questa tesi è progettare un sistema di controllo per una specifica classe di sistemi non lineari chiamati differentially flat, la quale include i droni quadrirotori. Il sistema di controllo mira a raggiungere prestazioni di tracciamento soddisfacenti sfruttando un processo Gaussiano come strumento di learning. La novità introdotta da questo lavoro rispetto alla letteratura è considerare esplicitamente l’incertezza degli input del processo Gaussiano e valutare come questa danneggi l’identificazione e il tracciamento. Inoltre, la tesi include un nuovo approccio per affrontare un problema aperto nella teoria del controllo: la robustezza dei sistemi learning-based. Infine, il sistema di controllo viene testato in simulazione sulla dinamica laterale di un drone quadrirotore. I risultati ottenuti dimostrano l’efficacia del sistema di controllo nell’ottenere un’ottima performance di tracciamento nonostante le limitazioni portate dall’incertezza degli input.
Learning-based control for UAV trajectory tracking
Sala, Greta
2020/2021
Abstract
Over the last decades, autonomous vehicle technologies have been advancing steadily. The versatility, the technological success and the cost reduction led the spread of drones and other self-driving vehicles in many research fields. Among the main functionalities that autonomous vehicles need are environmental perception, path planning and especially trajectory tracking. The capability to effectively follow a reference trajectory is critical to the successful completion of an autonomous vehicle’s mission. Nonetheless, the lack of accurate dynamic models prevents the design of control systems ensuring good tracking performance. Specifically, Unmanned Aerial Vehicles (UAVs), commonly known as drones, have nonlinear, dynamically coupled, and difficult to establish models. In recent times, researchers have resorted to machine learning algorithms to identify dynamical systems with limited prior knowledge on the model structure. Above all learning tools, Bayesian nonparametric modeling has shown strong potential for its flexibility and its efficiency in integrating available information on the model. In specific, Gaussian processes are Bayesian nonparametric approaches that have gained huge popularity because they can provide uncertainty estimates for their predictions. The goal of this thesis is to design a control system targeted for a specific class of nonlinear systems called differentially flat systems (including quadrotor UAVs). The control system aims at achieving satisfactory tracking performance by exploiting a Gaussian process as a learning tool. As a novelty compared to the literature on the subject, this work explicitly considers the inaccuracy of the Gaussian process inputs and evaluates how it damages identification and trajectory tracking. In addition, the thesis includes a novel approach to address an open problem in control theory: the robustness of learning-based systems. In conclusion, the control system is tested in simulation on the lateral dynamics of a quadcopter. The obtained results demonstrate the effectiveness of the control system in achieving good tracking despite the limitations brought by the input uncertainty.File | Dimensione | Formato | |
---|---|---|---|
2021_12_Sala.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary e Tesi
Dimensione
5.48 MB
Formato
Adobe PDF
|
5.48 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/183454