Author ORCID Identifier

https://orcid.org/0000-0001-7868-1756

Semester

Fall

Date of Graduation

2022

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Yu Gu

Committee Co-Chair

Jason Gross

Committee Member

Jason Gross

Committee Member

Guilherme Pereira

Committee Member

Natalia Schmid

Committee Member

Gianfranco Doretto

Abstract

Existing methods of autonomous robotic decision-making are often fragile when faced with inaccurate or incompletely modeled distributions of uncertainty, also known as ambiguity. While decision-making under ambiguity is a field of study that has been gaining interest, many existing methods tend to be computationally challenging, require many assumptions about the nature of the problem, and often require much prior knowledge. Therefore, they do not scale well to complex real-world problems where fulfilling all of these requirements is often impractical if not impossible. The research described in this dissertation investigates novel approaches to robotic decision-making strategies which are resilient to ambiguity that are not subject to as many of these requirements as most existing methods. The novel frameworks described in this research incorporate physical feedback, diversity, and swarm local interactions, three factors that are hypothesized to be key in creating resilience to ambiguity. These three factors are inspired by examples of robots which demonstrate resilience to ambiguity, ranging from simple vibrobots to decentralized robotic swarms. The proposed decision-making methods, based around a proposed framework known as Ambiguity Trial and Error (AT&E), are tested for both single robots and robotic swarms in several simulated robotic foraging case studies, and a real-world robotic foraging experiment. A novel method for transferring swarm resilience properties back to single agent decision-making is also explored. The results from the case studies show that the proposed methods demonstrate resilience to varying types of ambiguities, both stationary and non-stationary, while not requiring accurate modeling and assumptions, large amounts of prior training data, or computationally expensive decision-making policy solvers. Conclusions about these novel methods are then drawn from the simulation and experiment results and the future research directions leveraging the lessons learned from this research are discussed.

Download

Included in

Robotics Commons

Share

COinS