Multi-Robot Systems (MRS) are gaining increasing attention in the research community due to their broad range of potential applications and the numerous advantages they offer over single-robot systems. These advantages include improved scalability, robustness to individual failures, and faster task completion. However, the deployment of MRS in real-world environments is considerably more challenging, as it requires effective coordination and cooperation among multiple agents to achieve common objectives. These challenges become even more demanding under non-ideal conditions, such as limited communication, sensing uncertainties, and dynamic environments, where truly decentralized and reliable solutions are still lacking. In this thesis, we propose novel control strategies aimed at advancing Multi-Robot Systems toward practical real-world deployment and widespread adoption, with a particular focus on operating under uncertainties and incomplete information. First, we focus on the decentralized control of a team of robots equipped with limited-range sensing capabilities. Building upon traditional coverage control frameworks, we develop strategies for environment exploration and monitoring that enhance the robots' autonomy and efficiency when operating in large-scale environments. Furthermore, we extend the coverage control paradigm to human-robot interaction, enabling robots to adapt their motion based on human intentions and to safely navigate environments shared with humans, also exploring mixed reality solutions to facilitate this interaction. Taking a step further, we investigate even more challenging operating conditions, focusing on scenarios in which robots are equipped solely with camera-like sensors featuring a limited angular field of view for perceiving their teammates. We first consider strategies that maintain other robots within each agent’s visual field to ensure continuous information exchange. Building on this idea, we then propose alternative approaches suited for larger teams, where maintaining visual contact with all other robots is no longer feasible. Among these, we develop an exploration–exploitation strategy that balances neighbor seeking and task execution, as well as a clustering framework that partitions the team into subgroups and coordinates each subgroup as a single agent. Finally, we explore learning-based solutions, which offer significant advantages by enabling robots to autonomously adapt to complex, uncertain, and dynamic environments without requiring explicit modeling of all system dynamics or interactions. In particular, we investigate decentralized coverage control using both reinforcement learning and neural networks, allowing each robot to learn an effective control policy based solely on locally available information. Furthermore, we employ Gaussian Processes to reconstruct spatial phenomena under limited sensing conditions, enabling the robots to collectively estimate and monitor environmental distributions with improved accuracy. In conclusion, we address the problem of online learning of uncertainties in a robot's motion model, with the goal of extending this capability to multi-robot scenarios such as close-formation control under disturbances (e.g., downwash), cooperative transportation, and other collaborative tasks.
I Sistemi Multi-Robot (MRS) stanno attirando un’attenzione crescente nella comunità scientifica grazie alla loro vasta gamma di potenziali applicazioni e ai numerosi vantaggi che offrono rispetto ai sistemi a singolo robot. Tra questi vantaggi si annoverano una maggiore scalabilità, una migliore robustezza ai guasti dei singoli agenti e una più rapida esecuzione dei compiti. Tuttavia, l’impiego dei MRS in ambienti reali risulta notevolmente più complesso, poiché richiede un’efficace coordinazione e cooperazione tra molteplici agenti per raggiungere obiettivi comuni. Queste sfide diventano ancora più impegnative in condizioni non ideali, come comunicazione limitata, incertezze sensoriali e ambienti dinamici, dove soluzioni realmente decentralizzate e affidabili sono ancora carenti. In questa tesi, proponiamo nuove strategie di controllo volte a favorire l’avanzamento dei Sistemi Multi-Robot verso un impiego pratico nel mondo reale e una diffusione su larga scala, con particolare attenzione al funzionamento in presenza di incertezze e informazioni incomplete. Inizialmente, ci concentriamo sul controllo decentralizzato di un team di robot dotati di sensori a raggio limitato. Basandoci sui tradizionali framework di coverage control, sviluppiamo strategie di esplorazione e monitoraggio dell’ambiente che migliorano l’autonomia e l’efficienza dei robot durante il funzionamento in ambienti di grande scala. Inoltre, estendiamo il paradigma del coverage control all’interazione uomo-robot, permettendo ai robot di adattare il proprio movimento in base alle intenzioni umane e di navigare in sicurezza in ambienti condivisi con le persone, esplorando anche soluzioni di Mixed Reality per facilitare tale interazione. Successivamente, affrontiamo condizioni operative ancora più complesse, concentrandoci su scenari in cui i robot sono equipaggiati esclusivamente con sensori di tipo visivo, come ad esempio telecamere, con un campo visivo angolare limitato per percepire i propri compagni. In un primo momento, consideriamo strategie che mantengano gli altri robot all’interno del campo visivo di ciascun agente, garantendo così uno scambio continuo di informazioni. Partendo da questa idea, proponiamo poi approcci alternativi adatti a team di dimensioni maggiori, nei quali mantenere il contatto visivo con tutti gli altri robot non è più fattibile. Tra questi, sviluppiamo una strategia di exploration–exploitation che bilancia la ricerca dei vicini con l’esecuzione della missione, oltre a un framework di clustering che suddivide il team in sottogruppi e coordina ciascun sottogruppo come un singolo agente. Infine, esploriamo soluzioni basate sull’apprendimento, che offrono vantaggi significativi consentendo ai robot di adattarsi autonomamente ad ambienti complessi, incerti e dinamici senza richiedere una modellazione esplicita di tutte le dinamiche o interazioni del sistema. In particolare, indaghiamo il coverage control decentralizzato utilizzando sia Reinforcement Learning sia reti neurali, permettendo a ciascun robot di apprendere una politica di controllo efficace basata esclusivamente sulle informazioni locali disponibili. Inoltre, impieghiamo Processi Gaussiani per ricostruire fenomeni spaziali in condizioni di utilizzo di sensori con portata limitata, consentendo ai robot di stimare e monitorare collettivamente processi ambientali con maggiore accuratezza. In conclusione, affrontiamo il problema dell’apprendimento online delle incertezze nel modello di movimento di un robot, con l’obiettivo di estendere tale capacità a scenari multi-robot come il controllo in formazione ravvicinata in presenza di disturbi (ad esempio il downwash), il trasporto cooperativo e altri compiti collaborativi.
Controllo e Cooperazione Autonoma in un Sistema Multi-Robot: Strategie Decentralizzate in Condizioni di Sensorizzazione Limitata e Incertezza
CATELLANI, MATTIA
2026
Abstract
Multi-Robot Systems (MRS) are gaining increasing attention in the research community due to their broad range of potential applications and the numerous advantages they offer over single-robot systems. These advantages include improved scalability, robustness to individual failures, and faster task completion. However, the deployment of MRS in real-world environments is considerably more challenging, as it requires effective coordination and cooperation among multiple agents to achieve common objectives. These challenges become even more demanding under non-ideal conditions, such as limited communication, sensing uncertainties, and dynamic environments, where truly decentralized and reliable solutions are still lacking. In this thesis, we propose novel control strategies aimed at advancing Multi-Robot Systems toward practical real-world deployment and widespread adoption, with a particular focus on operating under uncertainties and incomplete information. First, we focus on the decentralized control of a team of robots equipped with limited-range sensing capabilities. Building upon traditional coverage control frameworks, we develop strategies for environment exploration and monitoring that enhance the robots' autonomy and efficiency when operating in large-scale environments. Furthermore, we extend the coverage control paradigm to human-robot interaction, enabling robots to adapt their motion based on human intentions and to safely navigate environments shared with humans, also exploring mixed reality solutions to facilitate this interaction. Taking a step further, we investigate even more challenging operating conditions, focusing on scenarios in which robots are equipped solely with camera-like sensors featuring a limited angular field of view for perceiving their teammates. We first consider strategies that maintain other robots within each agent’s visual field to ensure continuous information exchange. Building on this idea, we then propose alternative approaches suited for larger teams, where maintaining visual contact with all other robots is no longer feasible. Among these, we develop an exploration–exploitation strategy that balances neighbor seeking and task execution, as well as a clustering framework that partitions the team into subgroups and coordinates each subgroup as a single agent. Finally, we explore learning-based solutions, which offer significant advantages by enabling robots to autonomously adapt to complex, uncertain, and dynamic environments without requiring explicit modeling of all system dynamics or interactions. In particular, we investigate decentralized coverage control using both reinforcement learning and neural networks, allowing each robot to learn an effective control policy based solely on locally available information. Furthermore, we employ Gaussian Processes to reconstruct spatial phenomena under limited sensing conditions, enabling the robots to collectively estimate and monitor environmental distributions with improved accuracy. In conclusion, we address the problem of online learning of uncertainties in a robot's motion model, with the goal of extending this capability to multi-robot scenarios such as close-formation control under disturbances (e.g., downwash), cooperative transportation, and other collaborative tasks.| File | Dimensione | Formato | |
|---|---|---|---|
|
Catellani.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
17.52 MB
Formato
Adobe PDF
|
17.52 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/362884
URN:NBN:IT:UNIMORE-362884