An autonomous robot is designed to execute tasks without direct human intervention. To do so, it requires some capabilities. Among them, one of the most challenging is the so-called Simultaneous Mapping And Localization (SLAM) problem. Solving the SLAM problem requires that the robot builds a map representing the environment, while also localizing itself within the map. The SLAM problem is a highly researched topic within the scientific community. Consequently, many algorithms have been developed to address the SLAM problem, and it is critical to determine which approach is best suited in given contexts. This requires metrics and approaches to evaluate SLAM algorithms based on their performance. Some methods perform an a posteriori evaluation i.e., after the process is over. Sometimes, this can be limiting, necessitating the use of approaches that aim for a priori evaluation, estimating the performance of a robot (e.g., the error that it will make) before it has actually mapped the environment. Our thesis aims to compare different predictive models for making a priori predictions about the performance of a SLAM algorithm, which is modeled as a regression problem. Our approach implements methodologies to extract features from floor plans of environments. These features are used to train two learning models for predicting the SLAM algorithm’s performance, the linear model and Gaussian process model. The third model uses a convolutional neural network approach, which finds correlations between floor plan images and the performance of the SLAM algorithm in the environments. The three predictive models are compared and evaluated against different metrics, assessing their adaptability to the data and the accuracy shown.
Un robot autonomo è progettato per eseguire compiti senza l’intervento diretto dell’uomo. Per farlo, richiede alcune capacità. Tra queste, una delle più impegnative da realizzare è relativa al cosiddetto problema di Simultaneous Mapping And Localization (SLAM). La soluzione del problema SLAM richiede che il robot costruisca una mappa che rappresenta l’ambiente e che si localizzi all’interno della mappa. Il problema SLAM è un argomento molto studiato dalla comunità scientifica. Di conseguenza, sono stati sviluppati molti algoritmi per affrontare il problema SLAM ed è fondamentale determinare quale metodo sia più adatto in determinati contesti. Ciò richiede metriche e approcci per valutare gli algoritmi SLAM in base alle loro prestazioni. Alcuni metodi eseguono una valutazione a posteriori, cioè al termine del processo di scansione. A volte questo può essere limitante, rendendo necessario l’uso di approcci che mirano a una valutazione a priori, per esempio stimando l’errore che il robot commetterà nella costruzione della mappa prima di aver effettivamente mappato l’ambiente. Il nostro lavoro di tesi mira a confrontare diversi modelli predittivi per fare previsioni a priori sulle prestazioni di un algoritmo SLAM, che viene modellato come un problema di regressione. Il nostro approccio estrae caratteristi che dalle planimetrie degli ambienti. Queste caratteristiche sono utilizzate per addestrare due modelli di apprendimento per la regressione delle prestazioni dell’algoritmo SLAM, il modello lineare e il modello Gaussian process. Il terzo modello utilizza un approccio basato su una rete neurale convoluzionale, che trova correlazioni tra le immagini delle planimetrie e le prestazioni dell’algoritmo SLAM negli ambienti. I tre modelli predittivi sono confrontati e valutati in base a diverse metriche, valutando la loro adattabilità ai dati e l’accuratezza dimostrata.
Comparing models for predicting performance of SLAM algorithms
Capponcelli, Angelo
2023/2024
Abstract
An autonomous robot is designed to execute tasks without direct human intervention. To do so, it requires some capabilities. Among them, one of the most challenging is the so-called Simultaneous Mapping And Localization (SLAM) problem. Solving the SLAM problem requires that the robot builds a map representing the environment, while also localizing itself within the map. The SLAM problem is a highly researched topic within the scientific community. Consequently, many algorithms have been developed to address the SLAM problem, and it is critical to determine which approach is best suited in given contexts. This requires metrics and approaches to evaluate SLAM algorithms based on their performance. Some methods perform an a posteriori evaluation i.e., after the process is over. Sometimes, this can be limiting, necessitating the use of approaches that aim for a priori evaluation, estimating the performance of a robot (e.g., the error that it will make) before it has actually mapped the environment. Our thesis aims to compare different predictive models for making a priori predictions about the performance of a SLAM algorithm, which is modeled as a regression problem. Our approach implements methodologies to extract features from floor plans of environments. These features are used to train two learning models for predicting the SLAM algorithm’s performance, the linear model and Gaussian process model. The third model uses a convolutional neural network approach, which finds correlations between floor plan images and the performance of the SLAM algorithm in the environments. The three predictive models are compared and evaluated against different metrics, assessing their adaptability to the data and the accuracy shown.File | Dimensione | Formato | |
---|---|---|---|
2024_07_Capponcelli_Executive_Summary.pdf
accessibile in internet per tutti
Descrizione: testo executive summary
Dimensione
996.77 kB
Formato
Adobe PDF
|
996.77 kB | Adobe PDF | Visualizza/Apri |
2024_07_Capponcelli_Tesi.pdf
accessibile in internet per tutti
Descrizione: testo tesi
Dimensione
4.5 MB
Formato
Adobe PDF
|
4.5 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/10589/223090