Robots are increasingly involved in daily activities. Especially considering the Factory of the Future paradigm, humans and robots have to efficiently collaborate: the robot complements the human’s capabilities, learning new tasks and adapting itself to compensate for uncertainties. The need to provide intelligence to robotic systems is therefore more and more required. With this aim, this Thesis focuses on the investigation of machine learning techniques to make a robot able to learn and (re)optimize an industrial assembly task. Two main contributions are defined: (1) a trajectory learning algorithm (to teach to the robot the execution of the assembly task from human demonstrations), and (2) a (re)optimization procedure of the task execution (to optimally tune control parameters improving the assembly task performance). Considering (1), the most consistent reference trajectory (among a number of human-demonstrated ones) can be automatically selected to perform the assembly task. Considering (2), a Bayesian Optimization-based algorithm has been designed to (re)optimize the assembly task performance, making the robot able to compensate for task uncertainties (e.g., parts positioning). To validate the proposed methodology, an assembly task has been selected as an application. The task consists of mounting three shafts and the corresponding gears on a fixed base to simulate an assembly of a simple gearbox. The experiments, carried out on a population of 10 subjects, show the effectiveness of the proposed strategy, making the robot able to learn and (re)optimize its behaviour to accomplish the assembly task, even in the presence of task uncertainties.
La presenza di robot all’interno della società odierna è sempre più diffusa. In modo particolare, nella Fabbrica del Futuro, uomo e robot collaborano insieme: il robot completa le capacità dell’uomo, imparando nuovi task e adattandosi a compensare le incertezze. Da qui, la necessità di indagare i nuovi approcci di intelligenza artificiale che dovranno essere in grado di superare le sfide proposte da questo nuovo scenario. La presente Tesi di Laurea prende in considerazione il campo del machine learning applicato a manipolatori leggeri in contesto industriale, ricercando tecniche che consentano al robot di imparare e ottimizzare un task di assemblaggio. Due contributi principali concorrono a tale scopo: (1) un algoritmo per l’apprendimento di traiettorie (con l’intento di insegnare al robot l’esecuzione del task a partire da un set di dimostrazioni fatte dall’uomo), e (2) una procedura di (ri)ottimizzazione dell’esecuzione del task (che ricerca i parametri del controllo tali da ottimizzare la performance). In riferimento al punto (1), la traiettoria di riferimento è stata generata con una procedura automatica di selezione, come la più coerente tra quelle insegnate al robot da un operatore umano. Relativamente al punto (2), è stata scelta una strategia di ottimizzazione bayesiana con l’obiettivo di sviluppare una procedura di (ri)ottimizzazione per migliorare la performance dell’esecuzione del task. Per validare la metodologia proposta, un compito di assemblaggio è stato selezionato come applicazione target. Il task consiste nel montaggio di tre alberi e le corrispondenti ruote dentate, simulando l’assemblaggio di un riduttore semplificato. Gli esperimenti, effettuati su un campione di 10 soggetti, mostrano l’efficacia della procedura proposta nell’imparare e (ri)ottimizzare il task di assemblaggio, anche in presenza di incertezze.
Human-robot collaboration in assembly task learning enhanced by uncertainties adaptation via Bayesian optimization
MAGNI, MAURO;CANTONI, MARTINA
2018/2019
Abstract
Robots are increasingly involved in daily activities. Especially considering the Factory of the Future paradigm, humans and robots have to efficiently collaborate: the robot complements the human’s capabilities, learning new tasks and adapting itself to compensate for uncertainties. The need to provide intelligence to robotic systems is therefore more and more required. With this aim, this Thesis focuses on the investigation of machine learning techniques to make a robot able to learn and (re)optimize an industrial assembly task. Two main contributions are defined: (1) a trajectory learning algorithm (to teach to the robot the execution of the assembly task from human demonstrations), and (2) a (re)optimization procedure of the task execution (to optimally tune control parameters improving the assembly task performance). Considering (1), the most consistent reference trajectory (among a number of human-demonstrated ones) can be automatically selected to perform the assembly task. Considering (2), a Bayesian Optimization-based algorithm has been designed to (re)optimize the assembly task performance, making the robot able to compensate for task uncertainties (e.g., parts positioning). To validate the proposed methodology, an assembly task has been selected as an application. The task consists of mounting three shafts and the corresponding gears on a fixed base to simulate an assembly of a simple gearbox. The experiments, carried out on a population of 10 subjects, show the effectiveness of the proposed strategy, making the robot able to learn and (re)optimize its behaviour to accomplish the assembly task, even in the presence of task uncertainties.File | Dimensione | Formato | |
---|---|---|---|
2019_12_Magni_Cantoni.pdf
solo utenti autorizzati dal 03/12/2020
Descrizione: Testo della tesi
Dimensione
130.39 MB
Formato
Adobe PDF
|
130.39 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/10589/150736