Files in this item

FilesDescriptionFormat

application/pdf

application/pdfYOON-DISSERTATION-2019.pdf (3MB)
(no description provided)PDF

Description

Title:Path planning and control of flying robots with account of human’s safety perception
Author(s):Yoon, Hyung Jin
Director of Research:Hovakimyan, Naira
Doctoral Committee Chair(s):Hovakimyan, Naira
Doctoral Committee Member(s):Wang, Ranxiao F.; Stipanović, Dušan M.; Schwing, Alexander
Department / Program:Mechanical Sci & Engineering
Discipline:Mechanical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):Human-robot interaction
Optimal trajectory generation
Model predictive control
Reinforcement learning
Partially observable Markov decision process
Abstract:In this dissertation, a framework for planning and control of flying robot with the account of human’s safety perception is presented. The framework enables the flying robot to consider the human’s perceived safety in path planning. First, a data-driven model of the human’s safety perception is estimated from human’s test data using a virtual reality environment. A hidden Markov model (HMM) is considered for estimation of latent variables, as user’s attention, intention, and emotional state. Then, an optimal motion planner generates a trajectory, parameterized in Bernstein polynomials, which minimizes the cost related to the mission objectives while satisfying the constraints on the predicted human’s safety perception. Using Model Predictive Path Integral (MPPI) framework, the algorithm is possible to execute in real-time measuring the human’s spatial position and the changes in the environment. A HMM-based Q-learning is considered for computing the online optimal policy. The HMM-based Q-learning estimates the hidden state of the human in interactions with the robot. The state estimator in the HMM-based Q-learning infers the hidden states of the human based on past observations and actions. The convergence of the HMM-based Q-learning for a partially observable Markov decision process (POMDP) with finite state space is proved using stochastic approximation technique. As future research direction one can consider to use recurrent neural networks to estimate the hidden state in continuous state space. The analysis of the convergence of the HMM-based Q-learning algorithm suggests that the training of the recurrent neural network needs to consider both the state estimation accuracy and the optimality principle.
Issue Date:2019-04-16
Type:Text
URI:http://hdl.handle.net/2142/104824
Rights Information:Copyright 2019 Hyung Jin Yoon
Date Available in IDEALS:2019-08-23
Date Deposited:2019-05


This item appears in the following Collection(s)

Item Statistics