Files in this item

FilesDescriptionFormat

application/pdf

application/pdfDEKA-DISSERTATION-2019.pdf (6MB)
(no description provided)PDF

Description

Title:Reliable control of surgical robots and stabilization of long short-term memory neural networks
Author(s):Deka, Shankar A.
Director of Research:Stipanović, Dušan M.; Kesavadas, Thenkurussi
Doctoral Committee Chair(s):Hovakimyan, Naira
Doctoral Committee Member(s):Dullured, Geir E; Tomlin, Claire J
Department / Program:Mechanical Sci & Engineering
Discipline:Mechanical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):Controls
Dynamical systems
Neural networks
LSTM
stability
stabilization
Collision avoidance
Time-delays
Bilateral Teleoperation
Abstract:The work described in this thesis pertains to the design of safe and stabilizing controllers for robotic systems and neural networks. Problems such as collision avoidance in multi-arm robots, and safe control under time-delays in communication are formulated as a control system design problem and studied under a Lyapunov framework. A study on a widely used recurrent neural network is also presented from a dynamical systems viewpoint. In the first part of this thesis, we develop a strategy for collision avoidance among a system of robotic manipulators in the joint space. The resulting joint-feedback controller is obtained in a closed form, which means that the joint positions are directly used in computing the joint torques, without any additional intermediate steps for computing shortest distances or gradients of shortest distances between the links. Furthermore the collision avoidance controller can be augmented to any stable controller with different objectives, such as position tracking, velocity synchronization, coordination, formation control etc. As an example, set point stabilization is considered as a control objective and convergence of the joints to their desired positions is shown while guaranteeing collision avoidance among the links of the manipulators and avoiding deadlocks (unwanted local minima). The second part studies a well known problem in bilateral teleoperation systems in presence of asymmetric communication delays, for the case when there is a controller saturation constraint on both the local and remote manipulators. Stability under time delays has been a key challenge in the control of teleoperation systems, and several controllers, mostly based on the concept of passivity have been proposed over the past two decades. The recently proposed controllers with damping injection, improve tracking performance in addition to guaranteeing stability, due to explicit position information in the controller. We extend these results in the case of manipulators that can only produce bounded joint torques for coordination and force reflection. Sufficient conditions for guaranteeing stability and tracking are provided under such a controller restriction. The proposed controllers are validated using some simulation and experimental results. In the third part, we study some of the dynamical properties of Long Short-Term Memory Neural Networks (LSTMs). Such studies are needed in order have better insights into how and why LSTM models work so well with time-series data, with the ultimate goal of improving their training and performance. Towards that direction, two sufficient conditions on global asymptotic stability for autonomous LSTMs, are presented. One of these conditions is obtained in analytical form whereas the other is obtained through linear matrix inequality (LMI) based computation. Since these conditions are formulated in terms of the networks' weight matrices and biases that are essentially control variables, the same conditions can be viewed as a way to globally asymptotically stabilize these networks. These conditions and how to compute numerical values for the weight matrices and biases are illustrated by a number of numerical examples.
Issue Date:2019-08-14
Type:Text
URI:http://hdl.handle.net/2142/106138
Rights Information:Copyright 2019 Shankar Deka
Date Available in IDEALS:2020-03-02
Date Deposited:2019-12


This item appears in the following Collection(s)

Item Statistics