Among the different research fields in robotics, autonomous mobile robotics has been actively addressed in the last years. Autonomous exploration is one of the most important tasks that an autonomous robot, deployed in an initially unknown environment, must accomplish. The robot, with no prior information about the environment, has to choose where to move and consequently the best strategy to explore the environment in order to build its map incrementally. Over the years, different strategies have been proposed and developed. Even if classical techniques proved to be mostly successful, a recent research thrust aims to develop Machine Learning and in particular Deep Learning techniques to address the exploration problem, because of the performance achieved by these approaches in other fields. The purpose of this thesis is to compare classical and Deep Learning algorithms for exploration in order to understand what are the positive and negative sides of the different techniques. We compare them on different tasks and environments, obtaining comparable results for the classical and learning algorithms in most of the metrics considered. Results also highlight the difficulties faced by some learning algorithms when tested in more complex environments than those used in training.
La robotica mobile autonoma è stata molto studiata negli ultimi anni. Uno dei compiti più importanti che un robot autonomo, posto in un ambiente inizialmente sconosciuto, deve essere in grado di compiere è l'esplorazione. Il robot, senza alcuna conoscenza pregressa dell'ambiente, deve scegliere dove muoversi e conseguentemente la miglior strategia per esplorare l'ambiente stesso in modo da costruirne incrementalmente la mappa. Lungo gli anni diverse strategie sono state proposte e sviluppate. Anche se le tecniche classiche hanno dimostrato di avere successo nella maggior parte dei casi, una recente tendenza di ricerca mira a sviluppare algoritmi di Machine Learning e in particolare di Deep Learning per risolvere il problema dell'esplorazione, in virtù delle prestazioni raggiunte da queste tecniche in altri campi. Lo scopo di questa tesi è quello di confrontare algoritmi di esplorazione classici e algoritmi Deep Learning in modo da capire quali sono gli aspetti positivi e negativi delle differenti tecniche. Nella tesi abbiamo comparato questi algoritmi in base a compiti diversi e in ambienti diversi, ottenendo risultati simili per gli algoritmi classici e per quelli che fanno uso di learning per la maggior parte delle metriche considerate. I risultati ottenuti evidenziano anche le difficoltà incontrate da alcuni algoritmi che usano tecniche di learning quando vengono testati in ambienti più complessi rispetto a quelli di training.
Autonomous robot exploration using deep learning : an experimental analysis
PREMI, MARCO
2021/2022
Abstract
Among the different research fields in robotics, autonomous mobile robotics has been actively addressed in the last years. Autonomous exploration is one of the most important tasks that an autonomous robot, deployed in an initially unknown environment, must accomplish. The robot, with no prior information about the environment, has to choose where to move and consequently the best strategy to explore the environment in order to build its map incrementally. Over the years, different strategies have been proposed and developed. Even if classical techniques proved to be mostly successful, a recent research thrust aims to develop Machine Learning and in particular Deep Learning techniques to address the exploration problem, because of the performance achieved by these approaches in other fields. The purpose of this thesis is to compare classical and Deep Learning algorithms for exploration in order to understand what are the positive and negative sides of the different techniques. We compare them on different tasks and environments, obtaining comparable results for the classical and learning algorithms in most of the metrics considered. Results also highlight the difficulties faced by some learning algorithms when tested in more complex environments than those used in training.File | Dimensione | Formato | |
---|---|---|---|
Marco_Premi_Thesis.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
9.49 MB
Formato
Adobe PDF
|
9.49 MB | Adobe PDF | Visualizza/Apri |
Marco_Premi_Executive_summary.pdf
accessibile in internet per tutti
Descrizione: Executive summary
Dimensione
895.37 kB
Formato
Adobe PDF
|
895.37 kB | 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/10589/190127