A Learning-Based Framework for Quadrotor Modeling and Control - Monte Carlo Probabilistic Inference for Learning and COntrol (MC-PILCO) -研究报告

Aug 4, 2026

Upload by Anda

This thesis studies learning-based control of complex quadrotor dynamics using machine learning methods. It implements Model-Based Reinforcement Learning with the Monte Carlo Probabilistic Inference for Learning and COntrol (MC-PILCO) algorithm, an extension of PILCO that uses Gaussian Process Regression to learn quadrotor dynamics and assess suitability for flight control. The work models the quadrotor equations, simulates the learning environment, trains GPR with an RBF kernel and speed integration model, and evaluates Subset of Data to approximate the GP model. Results in simulation show improved prediction with more data, acceptable approximation with fewer points, and stronger control performance versus traditional controllers, supporting MC-PILCO as a viable alternative enabling future research.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages107

Reading Time4 h 30 min

AccessFree to download

Introduction

Quadrotor Model

Quadrotor Characteristic

Euler’s Angles

Quadrotor Dynamic Model

Quaternion Modeling

Quadrotor Control

Gaussian Processes for Regression

Weight-Space Model

Function-Space Model

Quadrotor Model Learning

Model Approximation

Quadrotor Control with MC-PILCO

Reinforcement Learning Framework

Master Thesis in Control Systems Engineering A Learning-Based Framework for Quadrotor Modeling and Control Master Candidate Mattia Scarpa Student ID 2005826 Supervisor Dott. Alberto Dalla Libera University of Padova Academic Year 2022/2023 In memory of my grandparents, Ada, Giuseppe, Mirella, and Alberto, to whom I dedicate this work! Abstract This work focuses on the study of controlling complex systems, such as quadrotors, through machine learning techniques. Speciﬁcally, it explores the implementation of the Model-Based Reinforcement Learning (MBRL) algorithm known as Monte Carlo Probabilistic Inference for Learning and COntrol (MCPILCO), an extension of the PILCO algorithm, which leverages Gaussian Process Regression (GPR) to accurately learn the dynamics of the quadrotor and investigates its potential applications in ﬂight control. The context of the research concerns the need to develop eﬀective and ﬂexible control methods for quadrotors, which must confront complex dynamics, model uncertainties, and operating environment uncertainties. In this respect, MC-PILCO, which relies on GPR, emerges as a promising technique for learning the system’s dynamics. The research involved studying the equations of quadrotor dynamics and using them to simulate the learning environment. The GPR model, used within the MC-PILCO algorithm, was trained using an RBF kernel and a speed integration model. Furthermore, the use of the Subset of Data method to approximate the GPR model was investigated, analyzing the eﬀect of reducing the number of data points on prediction accuracy. Experimental results in simulated environments show that the GPR model improves with an increase in data and that the Subset of Data method provides acceptable results within a certain degree of approximation. The eﬀectiveness of MC-PILCO was demonstrated by comparing its control results with those of other existing traditional controllers, showing substantial improvements. This demonstrates that the proposed MC-PILCO approach, which incorporates GPR, represents a valid alternative to the currently available simulation and control methods for quadrotors, paving the way for future research in this area. Sommario Questo elaborato si concentra sullo studio del controllo di sistemi complessi, come i quadrotor, attraverso l’uso di tecniche di apprendimento automatico. In particolare, viene esplorata l’implementazione dell’algoritmo di Model-Based Reinforcement Learning (MBRL) chiamato Monte Carlo Probabilistic Inference for Learning and COntrol (MC-PILCO), un’estensione dell’algoritmo PILCO, che sfrutta la Regressione del Processo Gaussiano (GPR) per apprendere accuratamente la dinamica del quadrotor e indagare le sue potenziali applicazioninel controllo del volo. Il contesto della ricerca riguarda la necessità di sviluppare metodi di controlloeﬃcaci e ﬂessibili per i quadrotor, i quali devono aﬀrontare dinamiche complesse, incertezze del modello e incertezze dell’ambiente operativo. A questo proposito, MC-PILCO, che si basa su GPR, emerge come una tecnica promettente per apprendere la dinamica del sistema. La ricerca ha comportato lo studio delle equazioni di dinamica del quadrotor e il loro utilizzo per simulare l’ambiente di apprendimento. Il modello GPR, utilizzato all’interno dell’algoritmo MC-PILCO, è stato addestrato utilizzando un kernel RBF e un modello di integrazione della velocità . Inoltre, è stata esaminatal’uso del metodo Subset of Data per approssimare il modello GPR, analizzandol’eﬀetto della riduzione del numero di punti dati sulla precisione della previsione. I risultati sperimentali, in ambienti simulati, mostrano che il modello GPRmigliora con un aumento dei dati e che il metodo Subset of Data forniscerisultatiaccettabili entro un certo grado di approssimazione. L’eﬃcacia del MC-PILCO è stata dimostrata confrontando i suoi risultati di controllo con quelli di altri controllori tradizionali già esistenti, mostrando m

FAQs

What problem does the thesis address in quadrotor control?

It targets the need for effective and flexible control methods for quadrotors under complex dynamics, model uncertainties, and operating-environment uncertainties.

How does MC-PILCO learn the quadrotor dynamics?

MC-PILCO uses Gaussian Process Regression (GPR) to learn system dynamics, trained with an RBF kernel and a speed integration model.

What is the role of the Subset of Data (SOD) method?

SOD approximates the GPR model, and the thesis studies how reducing the number of data points affects prediction accuracy.

Aug 4, 2026

Upload by Anda