Future science and exploration missions will implement innovative mission concepts to embark on daring endeavors exploiting cooperating intelligent systems. The concept of Distributed Space Systems boosts the achievable mission objectives due to the flexibility and adaptivity that the system inherently possesses. For instance, large synthetic instrumentation can be built by dedicated configurations of the system or very close-approach fly-by with small bodies can be performed; moreover, the single-point failure, typical of large satellites, is generally avoided in Distributed Systems and Formation Flying architecture. All these features are achieved at the cost of a significant increase in the on-board autonomy. Due to the relative distances involved, each agent composing the system needs to be able to rapidly and autonomously react to unforeseen events, such as collisions. To this end, it is critical that each spacecraft is capable of planning, navigating and controlling itself in unknown or partially-known environment, without ground-intervention. The research work presented here focuses on the development and testing of a full Guidance, Navigation & Control system aided by Artificial Intelligence techniques. The enhancement provided by Artificial Intelligence techniques allow the system to fly in uncertain environments by incrementally learning its mathematical modeling. The Thesis develops a number of methods to recover the underlying dynamics by simply measuring relative state and processing it with an Artificial Neural Network. This dynamics is then used to plan control action and enhance the navigation and control synthesis. In particular, three methods for dynamics reconstruction are developed, together with their mathematical foundation. The three approaches integrate the Artificial Neural Network at different levels: from fully integrated, where the dynamics is completely encapsulated into an Artificial Neural Network, to partially integrated in which the network learns either the unknown dynamical accelerations or reconstructs the unknown parameters of the analytical expression. Such reconstruction scheme is used in two different planning and control algorithms: Neural-Artificial Potential Field method and Model-Based Reinforcement Learning. The former is a fast and light algorithm that easily handles collision avoidance but lacks of planning; the latter is able to generate plans and control the spacecraft based on the learnt dynamics. Given the distributed planning architecture, each spacecraft does not know how the rest of the system is evolving. For this reason, it has been necessary to develop an AI-based routine coupling Long-Short Term Memory and Inverse Reinforcement Learning to predict the behavior of external agents, being either in free-motion or controlled-motion. The algorithms, when compared to classical methods, showed superior performance and constant increase in relevant Guidance, Navigation & control metrics (navigation accuracy, maneuvers Δv, etc.). Finally, in order to increase the Technology Readiness Level of the algorithms, the work presents the Processor-In-the-Loop testing campaign executed with relevant hardware: a micro-controller unit and a single-board computer with similar computational power with respect to flight-hardware. An end-to-end autocoding procedure has been developed to transition from Model-In-the-Loop simulations to Processor-In-the-Loop validation. The tests were deemed successful by evaluating the execution times, resource utilization and achieved accuracy. The outcome of the Thesis is a complete framework to integrate different AI-based techniques to enhance existing, well-established, algorithms. The methodology described here can easily be extended to other mission scenarios, where the flexibility and adaptivity of the system is critical.
Le missioni scientifiche e di esplorazione spaziale future implementeranno architetture innovative per intraprendere sfide inedite sfruttando sistemi intelligenti e cooperanti. L’idea dei Sistemi Spaziali Distribuiti incrementa il livello degli obiettivi raggiungibili grazie alla flessibilità e alla capacità di adattarsi che il sistema possiede intrinsecamente. Ad esempio, strumentazione, cosiddetta sintetica, di taglia considerevole potrebbe essere costruita sfruttando configurazioni dedicate del sistema oppure operazioni rischiose di prossimità ravvicinata attorno a corpi celesti sarebbero possibili grazie al minor costo e alla ridondanza del sistema; inoltre, Sistemi Spaziali Distribuiti, e il volo in formazione in generale, non presenterebbero un unico punto di fallimento catastrofico, caratteristica propria delle missioni che utilizzano satelliti monoliti. Tutti questi benefici sono raggiungibili al costo di un avanzamento considerevole dell’autonomia a bordo. A causa delle piccole distanze relative, ogni agente del sistema deve essere in grado di reagire prontamente ed autonomamente ad eventi inaspettati, per esempio ai rischi di collisioni. A tal fine, è importante che ciascun satellite sia in grado di pianificare, navigare e controllare sé stesso in ambienti anche poco noti, senza l’intervento da terra. La ricerca presentata in questo lavoro si focalizza sullo sviluppo e test di un sistema completo di Guida, Navigazione & Controllo accoppiato e supportato da tecniche di Intelligenza Artificiale. Il contributo garantito dall’Intelligenza Artificiale permette al sistema di volare in ambienti non perfettamente conosciuti incrementando la sua conoscenza dei modelli dinamici durante la missione. La Tesi sviluppa diversi metodi per ricostruire la dinamica partendo da semplici misure di stato relativo digerite da una Rete Neurale Artificiale. Tali dinamiche sono poi usate per pianificare le azioni di controllo e migliorare la navigazione e l’attuazione. In particolare, sono stati considerati tre metodi con i relativi fondamenti matematici. I tre diversi approcci integrano la rete neurale a differenti livelli: partendo dalla completa integrazione, in cui la dinamica è completamente rappresentata dalla rete, fino all’integrazione parziale in cui alla rete è affidato il ruolo di ricostruire parametri o accelerazioni non note supportando un modello analitico della dinamica del sistema. Tale procedura di ricostruzione dinamica è utilizzata in due algoritmi di Guida e Controllo: Potenziale Artificiale Neurale ad Apprendimento a Rinforzo basato su modello. Il primo è un algoritmo veloce e leggero che affronta il rischio di collisioni efficacemente ma non è in grado di pianificare e ottimizzare. Il secondo algoritmo è capace di produrre previsioni di controllo ottimale basati sulla dinamica ricostruita. Tuttavia, data l’architettura distribuita, ogni satellite non è a conoscenza dello stato del sistema e come esso si stia evolvendo. Per questo motivo è stato necessario sviluppare un metodo basato sull’Intelligenza Artificiale (Long-Short Term Memory e Apprendimento per Rinforzo Inverso) in grado di predire le traiettorie future degli agenti esterni, sia in regime di dinamica naturale o controllata. Gli algoritmi, paragonati ai metodi classici, hanno dimostrato performance superiori e un costante incremento nelle più importanti metriche di confronto nell’ambito della Guida, Navigazione & Controllo (accuratezza di navigazione, consumo di propellente, ecc.). Infine, al fine di incrementare il Livello di Prontezza Tecnologica degli algoritmi, il lavoro presenta la campagna di validazione utilizzando un processore in anello (PIL) simile quelli di volo. Una procedura completa di codifica è stata sviluppata per la transizione tra simulazioni numeriche e validazione con processore. I test hanno dimostrato risultati positivi in termini di utilizzo delle risorse, tempo di esecuzione e accuratezza. Il risultato globale della Tesi è una struttura per integrare diverse tecniche di Intelligenza Artificiale per migliorare gli algoritmi tradizionali. La metodologia descritta può essere facilmente adattata ad altri scenari di missione, dove la flessibilità ricopre un ruolo fondamentale.
AI-augmented guidance, navigation and control for proximity operations of distributed systems
Silvestrini, Stefano
2020/2021
Abstract
Future science and exploration missions will implement innovative mission concepts to embark on daring endeavors exploiting cooperating intelligent systems. The concept of Distributed Space Systems boosts the achievable mission objectives due to the flexibility and adaptivity that the system inherently possesses. For instance, large synthetic instrumentation can be built by dedicated configurations of the system or very close-approach fly-by with small bodies can be performed; moreover, the single-point failure, typical of large satellites, is generally avoided in Distributed Systems and Formation Flying architecture. All these features are achieved at the cost of a significant increase in the on-board autonomy. Due to the relative distances involved, each agent composing the system needs to be able to rapidly and autonomously react to unforeseen events, such as collisions. To this end, it is critical that each spacecraft is capable of planning, navigating and controlling itself in unknown or partially-known environment, without ground-intervention. The research work presented here focuses on the development and testing of a full Guidance, Navigation & Control system aided by Artificial Intelligence techniques. The enhancement provided by Artificial Intelligence techniques allow the system to fly in uncertain environments by incrementally learning its mathematical modeling. The Thesis develops a number of methods to recover the underlying dynamics by simply measuring relative state and processing it with an Artificial Neural Network. This dynamics is then used to plan control action and enhance the navigation and control synthesis. In particular, three methods for dynamics reconstruction are developed, together with their mathematical foundation. The three approaches integrate the Artificial Neural Network at different levels: from fully integrated, where the dynamics is completely encapsulated into an Artificial Neural Network, to partially integrated in which the network learns either the unknown dynamical accelerations or reconstructs the unknown parameters of the analytical expression. Such reconstruction scheme is used in two different planning and control algorithms: Neural-Artificial Potential Field method and Model-Based Reinforcement Learning. The former is a fast and light algorithm that easily handles collision avoidance but lacks of planning; the latter is able to generate plans and control the spacecraft based on the learnt dynamics. Given the distributed planning architecture, each spacecraft does not know how the rest of the system is evolving. For this reason, it has been necessary to develop an AI-based routine coupling Long-Short Term Memory and Inverse Reinforcement Learning to predict the behavior of external agents, being either in free-motion or controlled-motion. The algorithms, when compared to classical methods, showed superior performance and constant increase in relevant Guidance, Navigation & control metrics (navigation accuracy, maneuvers Δv, etc.). Finally, in order to increase the Technology Readiness Level of the algorithms, the work presents the Processor-In-the-Loop testing campaign executed with relevant hardware: a micro-controller unit and a single-board computer with similar computational power with respect to flight-hardware. An end-to-end autocoding procedure has been developed to transition from Model-In-the-Loop simulations to Processor-In-the-Loop validation. The tests were deemed successful by evaluating the execution times, resource utilization and achieved accuracy. The outcome of the Thesis is a complete framework to integrate different AI-based techniques to enhance existing, well-established, algorithms. The methodology described here can easily be extended to other mission scenarios, where the flexibility and adaptivity of the system is critical.| File | Dimensione | Formato | |
|---|---|---|---|
|
phdthesis_silvestrini.pdf
accessibile in internet per tutti
Descrizione: Tesi Dottorato
Dimensione
10.98 MB
Formato
Adobe PDF
|
10.98 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/10589/169603