Autonomous system safety: Towards targeted and scalable approaches for validation and behaviour analysis

Autonomous systems are rapidly making their way into the physical domain, where they have to interact with human users in unstructured and safety-critical scenarios. The complexity of the jobs autonomy is tasked with is reflected in the diversity of components that make up the software and hardware stacks of autonomous systems. This diversity enables the performance we see, but also poses a challenge when trying to analyze the safety and behaviour properties of the systems — preventing the feasibility of a one-size-fits-all approach. In this dissertation, we approach the problem of safety analysis through the development of targeted and scalable methodologies that can cater to various forms of autonomous system design. In the first part of this work, we introduce methods for the validation of deterministic and stochastic white box models. Using reachability analysis, we develop a real-time safety monitor for automated systems interacting with human agents. The monitor incorporates human motion prediction with data driven reachability to provide a statement of safety assurance with formal guarantees in an online manner. In the stochastic setting, we develop safety analysis methods through examining the exit-time distribution of dynamics governed by stochastic differential equations (SDEs). Given some appropriately defined SDE, the moments of exit-times from a safe set are computed through an infinite dimensional convex optimization, where the constraints consist of linear and PSD matrices, resulting in a semidefinite program (SDP). In the latter sections, we present methods for the validation and behaviour analysis of black and semi-black box systems. In particular, the problem of failure search is posed as a sequential decision making process, enabling the use of efficient reinforcement learning (RL) methods to uncover the failure space of the autonomy under test. Using similar adaptive search techniques, we address the comparison of autonomous agent policies through contrastive summaries. The structured search of behaviour summaries demonstrates both computational efficiency and high interpretability.

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

• Du, Peter Boyu

Contributors dc:contributor

• Driggs-Campbell, Katherine
• Dullerud, Geir
• Mitra, Sayan
• Ornik, Melkior
• Dong, Roy

• Autonomous Systems
• Safety
• Validation
• Machine Learning
• Robotics

Statement dc:rights

• Copyright 2023 Peter Du

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121525

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121525