In the context of autonomous driving, given the absence of the driver, the problem of path tracking takes on a fundamental role. Path tracking refers to the ability of a vehicle to precisely follow an assigned trajectory. Currently, there is a limited number of studies with experimental validation of path tracking control algorithms, in particular those related to learning based algorithms. The aim of this thesis work is to develop Feedforward-Feedback, Linear Quadratic Regulator, and Deep Reinforcement Learning algorithms for path tracking and validate these algorithms experimentally on the QCars (i.e., scaled fully autonomous vehicles produced by the Canadian company Quanser). The real tests conducted in laboratory are based on different reference trajectories, while several Key Performance Indicators are used to evaluate the results of the defined algorithms. By comparing the experimental results, the strengths and weaknesses of each developed algorithm are highlighted. Summing up, this work provides a thorough overview of control challenges and capabilities for path tracking in autonomous driving, highlighting the most promising approaches.
Nel contesto della guida autonoma, data l'assenza del conducente, il problema del path tracking assume un ruolo fondamentale. Per path tracking si intende la capacità di un veicolo di seguire con precisione una traiettoria assegnata. Attualmente, vi è un numero limitato di studi con validazione sperimentale degli algoritmi di controllo per il path tracking, in particolare quelli relativi agli algoritmi learning based. Lo scopo del presente lavoro di tesi è sviluppare algoritmi Feedforward-Feedback, Linear Quadratic Regulator e Deep Reinforcement Learning per il path tracking e validare tali algoritmi sperimentalmente sulle QCars (cioè, veicoli in scala completamente autonomi prodotti dall'azienda canadese Quanser). Le prove reali condotte in laboratorio si basano su diverse traiettorie di riferimento, mentre diversi Key Performance Indicators vengono utilizzati per valutare i risultati degli algoritmi definiti. Confrontando i risultati sperimentali, vengono evidenziati i punti di forza e di debolezza di ciascun algoritmo sviluppato. In sintesi, questo lavoro fornisce una panoramica approfondita delle sfide e delle capacità di controllo per il path tracking nella guida autonoma, mettendo in evidenza gli approcci più promettenti.
MODEL BASED AND DEEP REINFORCEMENT LEARNING PATH TRACKING CONTROL STRATEGIES EXPERIMENTALLY VALIDATED VIA SCALED FULLY AUTONOMOUS VEHICLES
OLIVIERI, IGNAZIO
2022/2023
Abstract
In the context of autonomous driving, given the absence of the driver, the problem of path tracking takes on a fundamental role. Path tracking refers to the ability of a vehicle to precisely follow an assigned trajectory. Currently, there is a limited number of studies with experimental validation of path tracking control algorithms, in particular those related to learning based algorithms. The aim of this thesis work is to develop Feedforward-Feedback, Linear Quadratic Regulator, and Deep Reinforcement Learning algorithms for path tracking and validate these algorithms experimentally on the QCars (i.e., scaled fully autonomous vehicles produced by the Canadian company Quanser). The real tests conducted in laboratory are based on different reference trajectories, while several Key Performance Indicators are used to evaluate the results of the defined algorithms. By comparing the experimental results, the strengths and weaknesses of each developed algorithm are highlighted. Summing up, this work provides a thorough overview of control challenges and capabilities for path tracking in autonomous driving, highlighting the most promising approaches.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Ignazio_Olivieri_586376.pdf
File consultabile da postazioni biblioteca
Descrizione: Documento Tesi di Laurea
Dimensione
2.47 MB
Formato
Adobe PDF
|
2.47 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/84394