The lack of know-how in the robotics field and the need to constantly adjust to the ever-changing market's demands are for sure the main bottlenecks in allowing the robot's spread on the majority of the factory's shop floors. The deployment of robots on the plants' production lines also requires robotic agents capable of understanding the environment in which they operate and able to promptly react to unexpected events. These needs are perfectly fulfilled by the employed Learning Semantic Behavior Trees from Demonstration method. This work proposes to exploit the convenience and intuitiveness of Learning from Demonstration (LfD) techniques to allow non-experts to quickly reprogram robots, endowing them with the Skills necessary to address the most diverse robotic Tasks. The understanding of the robotic agent's surroundings and the flexibility to adapt to unforeseen scenarios, instead, are tackled through the Semantic Behavior Trees (BTs) paradigm. The latter encapsulate both the Semantic meaning of the operations to perform and the sequence of robotic actions that contribute to Task completion. Such a sequence of actions is learnt leveraging a single Kinaesthetic demonstration in which the operator physically moves the robot in its reachable workspace. Among the operations the robot is expected to perform, this dissertation focuses on those where the exertion of a controlled force profile is essential. This is the context in which this work claims innovation as it is the only one able to address force-relevant robotic applications exploiting an LfD - Semantic BTs methodology. In particular, this dissertation exhibits a method remarkably capable of learning all the Hybrid Position-Force control parameters (the Constraint Frame, the Selection Matrix and the setpoints) from a single demonstration. The set of actions the robot has to carry out to achieve a certain Task completion is first devised by a PDDL Planner and later encapsulated in a Semantic BT which supervises the robotic execution. Meanwhile, a knowledge base monitors the scene and detects all the Semantic changes that occur in the robot's surroundings.
La mancanza del know-how nel campo della robotica e il bisogno di adattarsi continuamente alle richieste del mercato, sono sicuramente le principali avversità che scoraggiano la diffusione dei robot all’interno di determinate aziende. L’utilizzo dei manipolatori robotici nelle varie linee di produzione aziendale implica che essi siano in grado di comprendere l’ambiente nel quale sono inseriti a operare, anche al fine di reagire prontamente a possibili eventi inaspettati. Tali necessità giustificano l’utilizzo del metodo Learning Semantic Behavior Trees from Demonstration, il quale sfrutta la convenienza e l’intuitività delle tecniche relative a Learning from Demonstration (LfD) per permettere a operatori non esperti di poter programmare rapidamente i robot, dotandoli delle Skills necessarie per compiere le più svariate Tasks. La comprensione dell’ambiente circostante al manipolatore robotico e la flessibilità nell’adattamento a scenari inaspettati, invece, sono risolti mediante il paradigma dei Semantic Behavior Trees (BTs). Questi ultimi, racchiudono sia il significato Semantico delle operazioni da svolgere, sia le azioni necessarie per a portare al termine il Task. Tali azioni sono apprese avvalendosi di una singola dimostrazione nella quale l’operatore prende e muove il robot nello spazio limitrofo. L'elaborato si focalizza sulle operazioni in cui il robot è chiamato a esercitare una certa forza sugli oggetti presenti nell'ambiente. L'approccio presentato è in grado di affrontare la necessità di controllare le forze da esercitare, con una metodologia LfD - Semantic BTs. In particolare, la tesi presenta un metodo in grado di memorizzare tutti i parametri del controllo Ibrido Posizione-Forza (la Terna di Vincolo, la Matrice di Selezione e i setpoint) a partire da una singola dimostrazione. La sequenza di azioni che il robot deve compiere è, dapprima progettata da un PDDL Planner e, in seguito supervisionata da un Semantic BT, fintanto che tutte le azioni sono giunte a compimento. Nel frattempo, il knowledge base monitora la scena e identifica tutti i cambiamenti Semantici che hanno luogo nelle vicinanze del robot.
Learning semantic behavior trees through a single kinaesthetic demonstration for force-relevant applications
MALAVENDA, MATTEO
2022/2023
Abstract
The lack of know-how in the robotics field and the need to constantly adjust to the ever-changing market's demands are for sure the main bottlenecks in allowing the robot's spread on the majority of the factory's shop floors. The deployment of robots on the plants' production lines also requires robotic agents capable of understanding the environment in which they operate and able to promptly react to unexpected events. These needs are perfectly fulfilled by the employed Learning Semantic Behavior Trees from Demonstration method. This work proposes to exploit the convenience and intuitiveness of Learning from Demonstration (LfD) techniques to allow non-experts to quickly reprogram robots, endowing them with the Skills necessary to address the most diverse robotic Tasks. The understanding of the robotic agent's surroundings and the flexibility to adapt to unforeseen scenarios, instead, are tackled through the Semantic Behavior Trees (BTs) paradigm. The latter encapsulate both the Semantic meaning of the operations to perform and the sequence of robotic actions that contribute to Task completion. Such a sequence of actions is learnt leveraging a single Kinaesthetic demonstration in which the operator physically moves the robot in its reachable workspace. Among the operations the robot is expected to perform, this dissertation focuses on those where the exertion of a controlled force profile is essential. This is the context in which this work claims innovation as it is the only one able to address force-relevant robotic applications exploiting an LfD - Semantic BTs methodology. In particular, this dissertation exhibits a method remarkably capable of learning all the Hybrid Position-Force control parameters (the Constraint Frame, the Selection Matrix and the setpoints) from a single demonstration. The set of actions the robot has to carry out to achieve a certain Task completion is first devised by a PDDL Planner and later encapsulated in a Semantic BT which supervises the robotic execution. Meanwhile, a knowledge base monitors the scene and detects all the Semantic changes that occur in the robot's surroundings.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Malavenda_Tesi_01.pdf
solo utenti autorizzati dal 19/03/2025
Descrizione: Testo della tesi
Dimensione
64.22 MB
Formato
Adobe PDF
|
64.22 MB | Adobe PDF | Visualizza/Apri |
2024_04_Malavenda_Executive Summary_02.pdf
solo utenti autorizzati dal 19/03/2025
Descrizione: Testo dell'executive summary
Dimensione
17.45 MB
Formato
Adobe PDF
|
17.45 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/10589/219800