Certifiable synthesis and analysis for autonomy: Data-driven and analytical techniques

Over the past few decades, progress in control theory, robotics, and machine learning has enabled autonomy across diverse application domains such as the chemical industry, mechanical manufacturing, and the aerospace industry. Despite these advances, the development and implementation of autonomous systems continue to confront numerous technical challenges. One particular issue is ensuring the certifiability of autonomous systems. As these systems enter more and more safety-critical applications, constructing certifiable systems becomes a crucial task for the community. In traditional industrial applications such as aircraft manufacturing, this problem has been extensively addressed with established certification standards. However, the same cannot be said for emerging autonomous systems interacting with complex environments, for instance, autonomous vehicles. Certifiability in these contexts remains a distant goal. In this dissertation, we focus on two types of problems, namely, the synthesis problem and the analysis problem. We combine data-driven approaches and analytical approaches to solve these two types of problems. The core contributions of this dissertation include (1) A formal definition of a general synthesis problem for temporal logic specifications. (2) A learning-based approach that incorporates contraction theory into machine learning to construct a tracking controller for a given dynamical system. Moreover, the tracking error of the synthesized controller is formally bounded. (3) An optimization-based path planner for signal temporal logic specifications. By combining the path planner and the tracking controller, we solve the synthesis problem defined in (1). (4) The notion of reachability functions and a tool, NeuReach, that can automatically construct a reachability function for a black-box system. (5) Demonstration of the proposed approaches on a perception-based control synthesis problem. For all the proposed approaches, we arm them with rigorous theoretical analysis. On the experimental side, we evaluate the proposed approaches on a variety of benchmarks in simulation. Moreover, we deploy some of the approaches on a quadcopter to complete a trajectory-tracking task and a safe landing task.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Sun, Dawei

Contributors dc:contributor

• Mitra, Sayan
• Dullerud, Geir E.
• Srikant, Rayadurgam
• Belabbas, Mohamed Ali

• Machine Learning
• Control Theory
• Safe Autonomy
• Robotics

Statement dc:rights

• Copyright 2023 Dawei Sun

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122050

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122050