Learning-Based Safe and Robust Control for Multi-Agent Systems Shuo Yang

Format:

Book

Thesis/Dissertation

Author/Creator:

Yang, Shuo, author.

Contributor:

University of Pennsylvania. Electrical and Systems Engineering., degree granting institution.

Language:

English

Subjects (All):

0364.

0771.

0800.

0984.

Local Subjects:

0364.

0771.

0800.

0984.

Physical Description:

1 electronic resource (179 pages)

Contained In:

Dissertations Abstracts International 87-07B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2025

Language Note:

English

Summary:

AI-enabled systems have become ubiquitous and integral to safety-critical domains, e.g., autonomous vehicles and aerial robotics. Despite promising empirical results, decision-making for critical systems incorporating AI components require careful consideration, as failures may have catastrophic consequences. One key challenge is that various uncertainties will inevitably arise from system limitations, black-box models, or environmental factors, and inaccurate estimation of intrinsic uncertainties or failure to account for other agents in the environment can lead to hazardous behaviors.In this dissertation, we study how to develop safe and robust learning-based control policies under various uncertainties. In particular, it explores how tools from statistics, game theory and formal methods can empower uncertainty quantification, adaptation to other agents, and robust policy synthesis. This dissertation explores two directions: 1) safe learning and control in multi-agent systems, and 2) safe perception-based control for robotic systems.The first three chapters will mainly focus on safe learning and control in multi-agent systems. We first show how to develop robust control strategies in safety-critical systems when encountering unknown other agents, which is achieved by approximating the Nash equilibrium policy profile. Then, we seek to learn a more adaptable policy through reinforcement learning while ensuring physical safety. We also provide an option on explicitly quantifying behavior uncertainty for other agents using conformal prediction, a distribution-free statistical tool for uncertainty quantification.The latter three chapters study how to synthesize safe perception-based control policy for robotic systems. We first address the problem where the perception system is noisy and we quantify the uncertainty using conformal prediction, then leverages the uncertainty to inform the probabilistic safety control design. Next, we generalize the task from simple point-to-point navigation to a broader and more complex tasks, expressed by signal temporal logic. We represent the environment as a neural signed distance field and synthesize a safe perception-based control policy through self-supervised learning. Finally, we scale the control policy to a visuomotor diffusion policy and propose a safe and efficient training framework for over-parameterized models

Notes:

Advisors: Pappas, George J. Committee members: Matni, Nikolai; Lee, Insup; Fainekos, Georgios; Lindemann, Lars

Source: Dissertations Abstracts International, Volume: 87-07, Section: B.

Ph.D. University of Pennsylvania 2025

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798276001685

Access Restriction:

Restricted for use by site license