Theses and Dissertations
Date of Award
12-2023
Document Type
Thesis
Degree Name
Master of Science in Engineering (MSE)
Department
Electrical Engineering
First Advisor
Dimah Dera
Second Advisor
Hasina Huq
Third Advisor
Wenjie Dong
Abstract
Autonomous mobile robots are essential in various domains such as industry, manufacturing and healthcare. Navigating autonomously and avoiding obstacles are crucial tasks that involve localizing the robot to explore and map unknown environments without prior knowledge. Simultaneous localization and mapping (SLAM) present significant challenges. In this study, we introduce a new approach to address robust navigation and mapping of robot actions using Bayesian Actor-Critic (A2C) reinforcement learning. The A2C framework combines policy-based and value-based learning by dividing the model into two components: (1) the policy model (Actor) determines the actions based on the state, and (2) the value model (Critic) evaluates whether the agent's action depending on its return value from being ahead or behind in an environment/game. This feedback guides the training process in which both models interact and optimize their outputs over time. To achieve robust exploration and collision-free navigation, we develop a Bayesian A2C model that generates robot actions and quantifies the uncertainty associated with these actions. Our approach incorporates Bayesian inference and optimizes the variational posterior distribution over the unknown model parameters using the evidence lower bound (ELBO) objective. We employ a first-order Taylor series approximation to estimate the mean and covariance of the variational distribution when passed through non-linear functions in the A2C model. The propagated covariance estimates the robot's action uncertainty at the output of the actor-network. Experimental results demonstrate the superior robustness of the proposed Bayesian A2C model when exploring environments with high levels of noise compared to deterministic alternatives. Furthermore, the proposed framework has the potential for various applications where robustness and uncertainty quantification are crucial, such as underwater robots, medical robots, micro-robots and drones.
Recommended Citation
Pedraza, Bryan Joseph, "Robust Uncertainty Estimation Framework in Deep Reinforcement Learning for Active SLAM" (2023). Theses and Dissertations. 1431.
https://scholarworks.utrgv.edu/etd/1431
Comments
Copyright 2023 Bryan J. Pedraza. All Rights Reserved.
https://go.openathens.net/redirector/utrgv.edu?url=https://www.proquest.com/dissertations-theses/robust-uncertainty-estimation-framework-deep/docview/2928443951/se-2?accountid=7119