Legged robots received an increased attention these past years because they are a promising technology for many applications in environments where wheeled robots are not suited. However, locomotion for legged robots is only a partially solved problem: these robots walk, but they don’t reach the same level of performances that are found in animals. Therefore, to solve in a meaningful way this problem, an easy source of inspiration is found in the solutions adopted by animals: nature has had millions of years to improve their movements. In this work we present a new method for bio-inspired mobile robotics.The aim is the study of Central Pattern Generators (CPG) and the introduction of Chaotic Neural Networks (CNN) to control the gait of a legged robot. Circuits inspired by Central Pattern Generators were already used to control robots, especially quadrupeds, but a few applications for bipedal robots also exist. Instead, chaotic neural networks were considered mainly as a machine learning method. The objective of this work is to audit which one of these two methods is the most suitable to control the locomotion of legged robots. In order to explain the features of both methods, a theoretical analysis was made. Moreover some heuristic to find the values for the parameters of a CNN was defined. A detailed study of the computational costs followed. To simulate a controller for legged robots a new simulator for CNNs was developed, whose implementation is detailed. Afterwards, applications of both CPGs and CNNs to locomotion are described: at first focused on quadruped and bipedal robots and then on data acquired from human subjects. Lastly, a discussion of the features of the two methods is presented. In this section the choices to make when deciding between them are highlighted. The proposed method was found to be effective and more practical than CPGs, however it was found that it suffers from some minor problems that can be attributed to both the novelty of the model and the features of the considered robots. These problems will surely be solved in the near future, with the main objective to develop a working controller for legged locomotion based on CNN.
I robot dotati di zampe hanno ricevuto molta attenzione negli ultimi anni, perché offrono una tecnologia promettente per varie applicazioni in ambienti dove i robot con ruote hanno difficoltà. Il problema della locomozione a zampe non è completamente risolto: attualmente questo tipo di robot cammina, ma non in modo particolarmente agile. Questo lavoro si inserisce nel contesto della robotica mobile bio-ispirata. Lo scopo è lo studio dei Central Pattern Generator (abbreviati con CPG) e l’introduzione delle Reti Neurali Caotiche (CNN, acronimo dall’inglese) al fine di controllare l’andatura di robot con zampe. Tra questi due modelli, i Central Pattern Generator sono già stati usati per controllare robot, principalmente quadrupedi,ma esistono anche alcune applicazioni per bipedi. Invece, le reti neurali caotiche sono state considerate più come metodi di machine learning. L’obiettivo di questa tesi è quindi il verificare quale di questi due metodi sia più adatto al movimento di robot con zampe. È stata effettuata una analisi di entrambi i metodi al fine di illustrare le caratteristiche principali. Inoltre sono state definite alcune euristiche per trovare dei parametri corretti per le CNN. È seguito uno studio più dettagliato dei costi computazionali. Per poter simulare un controllore per robot con zampe si è sviluppata una nuova libreria per la simulazione delle reti neurali caotiche, la cui implementazione è descritta in dettaglio. In seguito sono descritte delle applicazioni sia dei CPG sia delle CNN alla locomozione: inizialmente di robot quadrupedi e bipedi, e in seguito per approssimare dati acquisiti da soggetti umani. Infine, dopo alcuni esperimenti in ambiente simulato, sono discusse le caratteristiche principali dei due metodi, cercando di evidenziare le scelte da effettuare per preferire uno dei due. Il nuovo metodo proposto si è rivelato efficace e sicuramente più versatile dei CPG, nonostante alcuni minori problemi che derivano sia dalla novità del modello utilizzato sia dalle caratteristiche dei sistemi robotici considerati. Questi problemi saranno sicuramente risolti nel futuro prossimo, con l'obiettivo principale di sviluppare un controllore per robot a zampe basato su reti neurali caotiche.
Apprendimento di movimenti ritmici in robot a zampe con reti neurali caotiche
BANA, MATTEO
2014/2015
Abstract
Legged robots received an increased attention these past years because they are a promising technology for many applications in environments where wheeled robots are not suited. However, locomotion for legged robots is only a partially solved problem: these robots walk, but they don’t reach the same level of performances that are found in animals. Therefore, to solve in a meaningful way this problem, an easy source of inspiration is found in the solutions adopted by animals: nature has had millions of years to improve their movements. In this work we present a new method for bio-inspired mobile robotics.The aim is the study of Central Pattern Generators (CPG) and the introduction of Chaotic Neural Networks (CNN) to control the gait of a legged robot. Circuits inspired by Central Pattern Generators were already used to control robots, especially quadrupeds, but a few applications for bipedal robots also exist. Instead, chaotic neural networks were considered mainly as a machine learning method. The objective of this work is to audit which one of these two methods is the most suitable to control the locomotion of legged robots. In order to explain the features of both methods, a theoretical analysis was made. Moreover some heuristic to find the values for the parameters of a CNN was defined. A detailed study of the computational costs followed. To simulate a controller for legged robots a new simulator for CNNs was developed, whose implementation is detailed. Afterwards, applications of both CPGs and CNNs to locomotion are described: at first focused on quadruped and bipedal robots and then on data acquired from human subjects. Lastly, a discussion of the features of the two methods is presented. In this section the choices to make when deciding between them are highlighted. The proposed method was found to be effective and more practical than CPGs, however it was found that it suffers from some minor problems that can be attributed to both the novelty of the model and the features of the considered robots. These problems will surely be solved in the near future, with the main objective to develop a working controller for legged locomotion based on CNN.File | Dimensione | Formato | |
---|---|---|---|
2015_12_Bana.pdf
accessibile in internet per tutti
Descrizione: Testo della tesi
Dimensione
7.15 MB
Formato
Adobe PDF
|
7.15 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/10589/115196