From the moment it started developing its own self-conscience, mankind has always seen automation as a key to solve any type of process, from the simplest to the most complex. Based on this desire, starting from the 50’s of the last century, some researchers started laying down the foundations of what has nowadays become a fully-fledged discipline: Artificial Intelligence (AI). While still being at an early developing stage, in the last few years AI was capable of expressing all of its potential, also thanks to the interest and support from various application fields, such as robotics and automation. On the contrary, the construction and architecture fields have not demonstrated a real interest in the opportunities and potentials offered from this innovative discipline. The main goal of this thesis is to encourage the construction industry to apply these new operative methodologies in the context of Artificial Intelligence. Specifically, an autonomous system is developed with the aim of assembling discrete elements in three-dimensional geometries inside a constrained space defined by a designer. In the first part of the thesis, the state of the art regarding Artificial Intelligence is described, analyzing its most important theoretical aspects and applications in the building construction industry. This systematic approach has narrowed down the available options, identifying the Reinforcement Learning and Imitation Learning as the right strategies for the chosen problem. A second research step regards the choice of the developing environment among the different solutions available on the market. After an initial series of experiments adopting algorithmic programming platforms, resulting in only partially effective solutions, a more complete and effective solution has been found in the Unity3D developing platform. By making use of the Unity ML-Agents Toolkit, an open-source AI development package, and the added modeling flexibility granted from the development engine, a series of experiments have been conducted, composing the original contribution of the thesis. Specifically, two developing environments have been designed in which specific learning algorithms, such as the Proximal Policy Optimization and Behavioral Cloning, have been deployed. In the final chapters, from the analysis of the gathered data, it is demonstrated that in order to obtain relevant results from the proposed complex simulations, the training environments need to be simplified and eventually tested with different training algorithms, in order to guide the experiment towards more proficient results. The proposed work projects a future in which machine and man will collaborate not only in the decision-making processes typical of the design phase, but also in the manufacturing and on-site construction. By adopting an interdisciplinary approach and by employing resources from other fields - robotics for example - it’s possible to outline increasingly more complex and independent systems, that can be capable of designing and manufacturing buildings in every aspect, and, by scaling the process, be capable of designing the distribution of entire cities.
Dal momento in cui ha iniziato a prendere coscienza di sè, l'uomo ha sempre visto nell'automatizzazione una chiave per la risoluzione di processi di qualsiasi entità, dai più semplici ai più complessi. Sulla base di questa motivazione, a partire dagli anni '50 del secolo scorso, alcuni ricercatori hanno iniziato ad erigere le colonne portanti di quella che oggi è diventata una vera e propria disciplina: l'intelligenza artificiale (IA). Pur essendo ancora ad un livello embrionale, negli ultimi decenni l'IA ha potuto esprimere tutte le proprie potenzialità, grazie anche all'interesse dimostrato da vari ambiti di applicazione, quali - ad esempio - la robotica. Il mercato relativo all'architettura e all'ingegneria edile, però, non ha mai dato prova di essere realmente attratta dalle possibilità offerte da questa nuova disciplina. La tesi si pone come obiettivo principale, dunque, quello di stimolare il settore delle costruzioni ad adottare nuovi metodi operativi nell'ambito dell'intelligenza artificiale. Nel dettaglio, si dà vita ad un sistema autonomo che sia in grado di assemblare elementi discreti con l'intento di generare geometrie tridimensionali in un volume definito dal progettista. Nella prima parte della trattazione si delinea lo stato dell'arte riguardante l'intelligenza artificiale, i suoi principali sviluppi teorici e le applicazioni nel contesto dell'industria delle costruzioni. Questo lavoro sistematico ha permesso di individuare nel Reinforcement Learning e nell'Imitation Learning la strada da seguire, dando il via ad una seconda indagine, relativa alla scelta dell'ambiente di sviluppo. Dopo aver effettuato una serie di prove adoperando la programmazione algoritmica di alcune piattaforme, rivelatasi solamente in parte efficace allo scopo, la scelta è ricaduta sull'utilizzo del motore fisico e di calcolo più efficiente proposto da Unity3D. Usufruendo della presenza in rete dell'applicativo open-source Unity ML-Agents Toolkit legato all'apprendimento artificiale e grazie all'estrema flessibilità di modellazione degli ambienti di lavoro, è stato possibile eseguire una serie di sperimentazioni, presentate nel contributo originale della trattazione. In particolare, sono stati realizzati due ambienti nei quali si sono potuti applicare alcuni algoritmi sviluppati all'interno del plug-in, quali Proximal Policy Optimization e Behavioral Cloning. Nella parte conclusiva, grazie all'analisi eseguita sui dati raccolti, si dimostra che per portare a compimento simulazioni complesse come quelle adottate, gli ambienti necessitino di essere semplificati ed eventualmente testati con l'ausilio di algoritmi differenti, al fine di guidare la prova verso risultati più performanti. Il lavoro svolto prospetta un futuro dove macchina e uomo potranno collaborare tanto nei processi decisionali insiti nella progettazione, quanto nella fase prettamente più realizzativa e cantieristica. Con uno sforzo di unione interdisciplinare e con risorse provenienti da altri ambiti - ad esempio quello della robotica - è plausibile pensare a sistemi sempre più complessi e indipendenti, che possano realizzare interi organismi edilizi o che riescano, pensando su vasta scala, a progettare la distribuzione di intere città.
Architectural intelligence. Nuovi metodi operativi nell'ambito dell'intelligenza artificiale applicata all'industria delle costruzioni
FORTIS, ENRICO
2018/2019
Abstract
From the moment it started developing its own self-conscience, mankind has always seen automation as a key to solve any type of process, from the simplest to the most complex. Based on this desire, starting from the 50’s of the last century, some researchers started laying down the foundations of what has nowadays become a fully-fledged discipline: Artificial Intelligence (AI). While still being at an early developing stage, in the last few years AI was capable of expressing all of its potential, also thanks to the interest and support from various application fields, such as robotics and automation. On the contrary, the construction and architecture fields have not demonstrated a real interest in the opportunities and potentials offered from this innovative discipline. The main goal of this thesis is to encourage the construction industry to apply these new operative methodologies in the context of Artificial Intelligence. Specifically, an autonomous system is developed with the aim of assembling discrete elements in three-dimensional geometries inside a constrained space defined by a designer. In the first part of the thesis, the state of the art regarding Artificial Intelligence is described, analyzing its most important theoretical aspects and applications in the building construction industry. This systematic approach has narrowed down the available options, identifying the Reinforcement Learning and Imitation Learning as the right strategies for the chosen problem. A second research step regards the choice of the developing environment among the different solutions available on the market. After an initial series of experiments adopting algorithmic programming platforms, resulting in only partially effective solutions, a more complete and effective solution has been found in the Unity3D developing platform. By making use of the Unity ML-Agents Toolkit, an open-source AI development package, and the added modeling flexibility granted from the development engine, a series of experiments have been conducted, composing the original contribution of the thesis. Specifically, two developing environments have been designed in which specific learning algorithms, such as the Proximal Policy Optimization and Behavioral Cloning, have been deployed. In the final chapters, from the analysis of the gathered data, it is demonstrated that in order to obtain relevant results from the proposed complex simulations, the training environments need to be simplified and eventually tested with different training algorithms, in order to guide the experiment towards more proficient results. The proposed work projects a future in which machine and man will collaborate not only in the decision-making processes typical of the design phase, but also in the manufacturing and on-site construction. By adopting an interdisciplinary approach and by employing resources from other fields - robotics for example - it’s possible to outline increasingly more complex and independent systems, that can be capable of designing and manufacturing buildings in every aspect, and, by scaling the process, be capable of designing the distribution of entire cities.File | Dimensione | Formato | |
---|---|---|---|
Tesi.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
150.11 MB
Formato
Adobe PDF
|
150.11 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/10589/149596