Model-free Learning and Interaction-aware Control for Safe Autonomous Driving
Liu, Kaiwen
2023
Abstract
Autonomous vehicles technologies have greatly advanced in recent years, with promises of various benefits, including improved safety, efficient transportation, increased accessibility, etc. However, ensuring safety is one of the major challenges in bringing them into reality. On one hand, autonomous vehicles may need to share the road with other road users (including human-driven vehicles, cyclists, pedestrians, etc.). On the other hand, the manufacturers have to cope with multiple sources of variability in these vehicles due to part-to-part differences, aging, degradation, or even in-field modifications. This dissertation focuses on addressing these challenges by developing a behavior planner that is able to account for interaction with human drivers and model-free learning-based safety supervisors that are able to adapt to different systems or operating environments. This dissertation first presents the design of a game-theoretic interaction-aware behavior planner. Inspired by Stackelberg Competition, predictive models for human interactions are developed based on the Leader-Follower Game, and a decision-making framework is proposed that integrates game-theoretic predictions, online estimation of other driver's uncertain interactions and optimal control with explicit safety characterization. The proposed approach is applied to forced merging scenarios, where interaction and negotiation with other drivers are typically required. A comprehensive set of simulation-based case studies and validations on naturalistic driving dataset are presented, where the proposed approach demonstrates a high success rate. The dissertation then introduces two model-free learning algorithms to design safety supervisors suitable for non-safety critical systems and for safety critical systems. The design of the safety supervisor relies on the reference governor scheme, which is an add-on scheme to enforce pointwise-in-time state and control constraints. In non-safety critical control systems, where the violation of the constraints is not desirable but does not lead to severe consequences, such systems may initially operate with constraint violations and learn over time to avoid them through less aggressive maneuvering. In safety critical cases, in which constraint violation may lead to catastrophic consequences, systems will initially operate conservatively, and then improve their performance as they learn more about constraint boundaries and maneuvers that approach the constraint boundary. The results include developments and demonstrations of novel algorithms and supporting theory for autonomous online learning to operate systems safely and non-conservatively. Several applications are considered including power management of electric vehicles, rollover avoidance of ground vehicles, and tanker truck rollover avoidance under liquid sloshing effects.Deep Blue DOI
Subjects
Autonomous Vehicles Game Theory Machine Learning Multi-Agent Systems Constrained Control Safety-critical Systems
Types
Thesis
Metadata
Show full item recordCollections
Remediation of Harmful Language
The University of Michigan Library aims to describe library materials in a way that respects the people and communities who create, use, and are represented in our collections. Report harmful or offensive language in catalog records, finding aids, or elsewhere in our collections anonymously through our metadata feedback form. More information at Remediation of Harmful Language.
Accessibility
If you are unable to use this file in its current format, please select the Contact Us link and we can modify it to make it more accessible to you.