DTA

Archivio Digitale delle Tesi e degli elaborati finali elettronici

 

Tesi etd-01302018-145854

Tipo di tesi
Perfezionamento
Autore
ANSARI, YASMIN TAUQEER
URN
etd-01302018-145854
Titolo
"Development of learning-based control frameworks to enable high-precision positioning/tracking in continuum/soft robotic manipulators"
Settore scientifico disciplinare
ING-IND/34
Corso di studi
INGEGNERIA - Biorobotics (Marie Curie ITN fellows)
Commissione
relatore LASCHI, CECILIA
Parole chiave
  • Kinematic controllers
  • Maxwell-slip model
  • neural networks
  • reinforcement learning
Data inizio appello
30/06/2018;
Disponibilità
completa
Riassunto analitico
Continuum and soft robotic manipulators refer to a new generation of bio-inspired robotic systems that are appealing due to their compliance and dexterity. These systems manipulate unstructured environments and exhibit a high degree of safe human-robot interaction, thereby, demonstrating capacities well-beyond those of its rigid counterparts. This is the fundamental reason behind their rapid rise for three main areas of application- locomotion, whole-arm-manipulation/grasping, and positioning/tracking. On the downside, their biologically inspired design introduces a number of unprecedented control challenges that renders traditional control approaches ineffective, qualifying it as an active area of research.

The aim of the thesis is to enable these systems with high-precision positioning/tracking capabilities, which is dependent upon the development of real-time robust inverse kinematic controllers. This is non-trivial because: (i) the framework must account not only for the redundant degrees-of-freedom but more importantly the viscoelastic losses of overlapping elastic components; (ii) the framework should be scalable to the wide range of actuation technologies and actuator arrangements; (iii) the framework should account for external environmental constraints or internal faults that greatly affect the behavior of these systems to the large number of degrees-of-freedom.

As a first contribution, this thesis formalizes the key-concepts for developing control frameworks as a guideline for future applications and provides comprehensive insight into the advances made in this respect over the years. As a second contribution, this thesis introduces the tools and mechanisms necessary to develop closed-loop learning-based control frameworks with simultaneous resolution of redundancy and viscoelastic effects. The key-inspiration is drawn from equilibrium-point theory, which studies how primates elicit voluntary arm movements under conditions where viscoelastic effects are highly dominant. It postulates that joint-torques are not explicitly calculated by the central nervous system, but rather, generated through variations in muscle length that can be actively adjusted from two independent sources, neural commands and peripheral reflex loops. An equivalent generalized framework is developed for continuum and soft robotic manipulators which allows to parameterize the kinematic mapping between task-space variables and actuator-space variables through two independent variables. The underlying principle relies on generating spatially localized data with two active parameters. It is shown that this is mathematically equivalent to capturing underlying affine geometrical properties, which is desirable to correct for elastic deformations in order to improve positioning/tracking tasks. The proposed framework is investigated from the viewpoint of two learning-based paradigms: (i) reinforcement learning – it is computationally formalized as a hierarchical cooperative multiagent reinforcement learning framework with multi-objective optimization; (ii) supervised learning – it is computationally formalized as a novel bootstrapping mechanism that exploits spatially localized goal-directed movements to generate an appropriate database for learning. Extensive experimental studies in each paradigm highlight the effectiveness of the proposed mechanism to develop control architectures that can successfully address the challenging control task.
File