The study of human-machine interaction as a unique control system has been one of the first research interests in engineering, with almost a century of years since the first works. At the same time, it is a crucial aspect of the most recent technological developments in application fields concerning, for example, collaborative robotics and artificial intelligence. The cross-domain nature characterizing this field of study can cause difficulties in finding a guiding line that links motor control theory, modeling approaches of physiological control systems, and identifying human-machine general control models in manipulative tasks. For this reason, I chose to start this thesis work by analyzing state-of-the-art linear models, from the first crossover model defined in the frequency domain to the successive optimal control model, to end with models including more detailed descriptions of physiologic subsystems and biomechanics. The motivation behind this effort is to have a complete view of the linear models that could be easily handled both in the time domain and in the frequency domain by using the well-established methodology in the classical linear systems and control theory. Such model-based approaches aiming to characterize human behavior have been, as said, practically applied in a wide variety of scenarios. Among them, human-robot interaction is one of the most exciting, particularly in tasks where a continuous physical interaction between humans and the controlled plant is present. In this context, the human subject can adapt its control behavior to the external sensed dynamics. This capability significantly affects the control delay, making its characterization and prevision a crucial aspect to understand. I will address this topic in the third chapter of the thesis, where a linear modeling approach that uniquely describes human and robot control actions will be proposed and experimentally validated in a collaborative robotic task. Such manipulation task was performed by ten different healthy subjects with a collaborative low-payload robot. In Human-Robot interaction, the possibility of increasing the intelligence and adaptability of the controlled plant by imitating human control behavior has been an objective of many research efforts in the last decades. From classical control-theory human control models to modern machine learning, neural networks, and reinforcement learning paradigms, the common denominator is the effort to model complex nonlinear dynamics typical of human activity. This suggests that our analysis can't be limited to the linear models treated above but must proceed with nonlinear dynamics and the efforts that have been made to reproduce them. The fourth chapter investigates state-of-the-art nonlinear modeling techniques from the perspective of human control, considering the different physiological districts involved as the starting point and then proceeding with data-driven and model-based techniques able to describe higher cognitive processes such as decision-making and the creation of long-term strategy. In the first place, transport systems are presented as an alternative technological scenario in which the discussed techniques have been mainly applied with success recently. Successively, going back to human-robot interaction, I propose a novel nonlinear modeling technique able to predict human force generated during a cooperative task with a controlled robot. The proposed Narmax model was constructed using an artificial neural network as a nonlinear functional approximator and was trained on the same dataset as the one used to validate the previous linear model. The same human model was then tested online with different subjects and, most importantly, on an industrial high-payload robot. This was done to demonstrate that the obtained performance were not derived from data overfitting but that a good generalization capability characterizes the model. While a deep stability analysis of the system was not in the scope of this work, to avoid unstable and dangerous behavior, the robot has been controlled with an impedance control strategy characterized by a low stiffness value, resulting in compliant and safe movements. Moreover, the framework was controlled with a frequency higher than human motion by more than one order of magnitude. A further step forward in showing how the proposed modeling technique can be useful can be pointed out by exploiting the partial knowledge that I have of the system, with reference to robot control law and its dynamics, to gain knowledge of the system from raw data in a simple and fast way. The case of study was once again the system delay, which this time was extracted from the human model's output with the aid of simple approximated system identification techniques. This process allows the user to easily extract the delayed information without complicated data processing. Moreover, by analyzing the characteristics of human response during the proposed cooperative interaction with the controlled robot, a regular presence of peak values is evident, as a first reaction to the external forcing function. Such peak values represent the most important feature to be known by the robot to anticipate human action, rather than having to estimate the whole force response sample by sample. For this reason, Peak-to-Peak Dynamics have been exploited to obtain a reduced-order model that is able to forecast the peak of human response in a reliable way. The effort in this sense is considerable, given that such techniques have been used in the past with different kinds of chaotic systems characterized by known attractors, but not to Narmax models.
Lo studio dell'interazione uomo-macchina come sistema di controllo unico è stato uno dei primi interessi di ricerca in ingegneria, con quasi un secolo di anni dai primi lavori. Allo stesso tempo, costituisce un aspetto cruciale dei più recenti sviluppi tecnologici in campi applicativi riguardanti, ad esempio, la robotica collaborativa e l’intelligenza artificiale. La natura interdisciplinare che caratterizza questo campo di studi può causare difficoltà nel trovare una linea guida che colleghi la teoria del controllo motorio, gli approcci di modellazione dei sistemi di controllo fisiologico e l’identificazione di modelli di controllo generale uomo-macchina in compiti manipolativi. Per questo motivo, ho scelto di iniziare questo lavoro di tesi analizzando i modelli lineari allo stato dell'arte, dal primo modello di crossover definito nel dominio della frequenza al successivo modello di controllo ottimo, per finire con modelli comprendenti descrizioni più dettagliate delle caratteristiche fisiologiche sottosistemi e biomeccanica. La motivazione alla base di questo sforzo è quella di avere una visione completa dei modelli lineari che potrebbero essere facilmente gestiti sia nel dominio del tempo che nel dominio della frequenza utilizzando la metodologia consolidata nei sistemi lineari classici e nella teoria del controllo. Tali approcci basati su modelli volti a caratterizzare il comportamento umano sono stati, come detto, applicati praticamente in un'ampia varietà di scenari. Tra questi, l'interazione uomo-robot è una delle più entusiasmanti, in particolare nei compiti in cui è presente un'interazione fisica continua tra l'uomo e l'impianto controllato. In questo contesto, il soggetto umano può adattare il proprio comportamento di controllo alle dinamiche esterne percepite. Questa capacità influenza in modo significativo il ritardo del controllo, rendendone la caratterizzazione e la previsione un aspetto cruciale da comprendere. Affronterò questo argomento nel terzo capitolo della tesi, dove un approccio di modellazione lineare che descrive in modo univoco le azioni di controllo umano e robotico sarà proposto e validato sperimentalmente in un compito robotico collaborativo. Tale compito di manipolazione è stato eseguito da dieci diversi soggetti sani con un robot collaborativo a basso carico utile. Nell'interazione uomo-robot, la possibilità di aumentare l'intelligenza e l'adattabilità della pianta controllata imitando il comportamento di controllo umano è stato un obiettivo di molti sforzi di ricerca negli ultimi decenni. Dai modelli di controllo umano classici della teoria del controllo all’apprendimento automatico moderno, alle reti neurali e ai paradigmi di apprendimento per rinforzo, il denominatore comune è lo sforzo di modellare dinamiche non lineari complesse tipiche dell’attività umana. Ciò suggerisce che la nostra analisi non può limitarsi ai modelli lineari sopra trattati ma deve procedere con le dinamiche non lineari e gli sforzi che sono stati fatti per riprodurle. Il quarto capitolo indaga le tecniche di modellazione non lineare all'avanguardia dal punto di vista del controllo umano, considerando i diversi distretti fisiologici coinvolti come punto di partenza e procedendo poi con tecniche data-driven e model-based in grado di descrivere processi cognitivi superiori come come processo decisionale e creazione di una strategia a lungo termine. In primo luogo, i sistemi di trasporto vengono presentati come uno scenario tecnologico alternativo in cui le tecniche discusse sono state applicate con successo principalmente di recente. Successivamente, tornando all'interazione uomo-robot, propongo una nuova tecnica di modellazione non lineare in grado di prevedere la forza umana generata durante un compito cooperativo con un robot controllato. Il modello Narmax proposto è stato costruito utilizzando una rete neurale artificiale come approssimatore funzionale non lineare ed è stato addestrato sullo stesso set di dati utilizzato per convalidare il modello lineare precedente. Lo stesso modello umano è stato poi testato online con soggetti diversi e, soprattutto, su un robot industriale ad alto carico utile. Ciò è stato fatto per dimostrare che le prestazioni ottenute non derivano da un overfitting dei dati ma che una buona capacità di generalizzazione caratterizza il modello. Sebbene un'analisi approfondita della stabilità del sistema non rientrasse nello scopo di questo lavoro, per evitare comportamenti instabili e pericolosi, il robot è stato controllato con una strategia di controllo dell'impedenza caratterizzata da un basso valore di rigidità, risultando in movimenti conformi e sicuri. Inoltre, la struttura era controllata con una frequenza superiore al movimento umano di più di un ordine di grandezza. Un ulteriore passo avanti nel mostrare come la tecnica di modellazione proposta possa essere utile può essere sottolineato sfruttando la parziale conoscenza che ho del sistema, con riferimento alle leggi di controllo del robot e alla sua dinamica, per acquisire conoscenza del sistema da dati grezzi in un modo semplice e veloce. Il caso di studio è stato ancora una volta il ritardo del sistema, che questa volta è stato estratto dall'output del modello umano con l'aiuto di semplici tecniche di identificazione del sistema approssimate. Questo processo consente all'utente di estrarre facilmente le informazioni ritardate senza complicate elaborazioni dei dati. Inoltre, analizzando le caratteristiche della risposta umana durante la proposta interazione cooperativa con il robot controllato, è evidente una presenza regolare di valori di picco, come prima reazione alla funzione di forzante esterna. Tali valori di picco rappresentano la caratteristica più importante che il robot deve conoscere per anticipare l'azione umana, piuttosto che dover stimare l'intera risposta della forza campione per campione. Per questo motivo, la dinamica picco-picco è stata sfruttata per ottenere un modello di ordine ridotto in grado di prevedere in modo affidabile il picco della risposta umana. Lo sforzo in questo senso è notevole, dato che tali tecniche sono state utilizzate in passato con diversi tipi di sistemi caotici caratterizzati da attrattori noti, ma non con modelli Narmax.
Modelli di Controllo e Stima dell'Intenzione nel Contesto dell'Interazione Uomo-Robot / Scibilia, Adriano. - (2023 Nov 13).
Modelli di Controllo e Stima dell'Intenzione nel Contesto dell'Interazione Uomo-Robot
SCIBILIA, ADRIANO
2023-11-13
Abstract
The study of human-machine interaction as a unique control system has been one of the first research interests in engineering, with almost a century of years since the first works. At the same time, it is a crucial aspect of the most recent technological developments in application fields concerning, for example, collaborative robotics and artificial intelligence. The cross-domain nature characterizing this field of study can cause difficulties in finding a guiding line that links motor control theory, modeling approaches of physiological control systems, and identifying human-machine general control models in manipulative tasks. For this reason, I chose to start this thesis work by analyzing state-of-the-art linear models, from the first crossover model defined in the frequency domain to the successive optimal control model, to end with models including more detailed descriptions of physiologic subsystems and biomechanics. The motivation behind this effort is to have a complete view of the linear models that could be easily handled both in the time domain and in the frequency domain by using the well-established methodology in the classical linear systems and control theory. Such model-based approaches aiming to characterize human behavior have been, as said, practically applied in a wide variety of scenarios. Among them, human-robot interaction is one of the most exciting, particularly in tasks where a continuous physical interaction between humans and the controlled plant is present. In this context, the human subject can adapt its control behavior to the external sensed dynamics. This capability significantly affects the control delay, making its characterization and prevision a crucial aspect to understand. I will address this topic in the third chapter of the thesis, where a linear modeling approach that uniquely describes human and robot control actions will be proposed and experimentally validated in a collaborative robotic task. Such manipulation task was performed by ten different healthy subjects with a collaborative low-payload robot. In Human-Robot interaction, the possibility of increasing the intelligence and adaptability of the controlled plant by imitating human control behavior has been an objective of many research efforts in the last decades. From classical control-theory human control models to modern machine learning, neural networks, and reinforcement learning paradigms, the common denominator is the effort to model complex nonlinear dynamics typical of human activity. This suggests that our analysis can't be limited to the linear models treated above but must proceed with nonlinear dynamics and the efforts that have been made to reproduce them. The fourth chapter investigates state-of-the-art nonlinear modeling techniques from the perspective of human control, considering the different physiological districts involved as the starting point and then proceeding with data-driven and model-based techniques able to describe higher cognitive processes such as decision-making and the creation of long-term strategy. In the first place, transport systems are presented as an alternative technological scenario in which the discussed techniques have been mainly applied with success recently. Successively, going back to human-robot interaction, I propose a novel nonlinear modeling technique able to predict human force generated during a cooperative task with a controlled robot. The proposed Narmax model was constructed using an artificial neural network as a nonlinear functional approximator and was trained on the same dataset as the one used to validate the previous linear model. The same human model was then tested online with different subjects and, most importantly, on an industrial high-payload robot. This was done to demonstrate that the obtained performance were not derived from data overfitting but that a good generalization capability characterizes the model. While a deep stability analysis of the system was not in the scope of this work, to avoid unstable and dangerous behavior, the robot has been controlled with an impedance control strategy characterized by a low stiffness value, resulting in compliant and safe movements. Moreover, the framework was controlled with a frequency higher than human motion by more than one order of magnitude. A further step forward in showing how the proposed modeling technique can be useful can be pointed out by exploiting the partial knowledge that I have of the system, with reference to robot control law and its dynamics, to gain knowledge of the system from raw data in a simple and fast way. The case of study was once again the system delay, which this time was extracted from the human model's output with the aid of simple approximated system identification techniques. This process allows the user to easily extract the delayed information without complicated data processing. Moreover, by analyzing the characteristics of human response during the proposed cooperative interaction with the controlled robot, a regular presence of peak values is evident, as a first reaction to the external forcing function. Such peak values represent the most important feature to be known by the robot to anticipate human action, rather than having to estimate the whole force response sample by sample. For this reason, Peak-to-Peak Dynamics have been exploited to obtain a reduced-order model that is able to forecast the peak of human response in a reliable way. The effort in this sense is considerable, given that such techniques have been used in the past with different kinds of chaotic systems characterized by known attractors, but not to Narmax models.File | Dimensione | Formato | |
---|---|---|---|
PhD_Thesis.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Licenza:
PUBBLICO - Pubblico con Copyright
Dimensione
5.52 MB
Formato
Adobe PDF
|
5.52 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.