Game-Theoretic and Set-Based Methods for Safe Autonomous Vehicles on Shared Roads

Abstract

Autonomous vehicle (AV) technology promises safer, cleaner, and more efficient transportation, as well as improved mobility for the young, elderly, and disabled. One of the biggest challenges of AV technology is the development and high-confidence verification and validation (V&V) of decision and control systems for AVs to safely and effectively operate on roads shared with other road users (including human-driven vehicles). This dissertation investigates game-theoretic and set-based methods to address this challenge. Firstly, this dissertation presents two game-theoretic approaches to modeling the interactions among drivers/vehicles on shared roads. The first approach is based on the "level-k reasoning" human behavioral model and focuses on the representation of heterogeneous driving styles of real-world drivers. The second approach is based on a novel leader-follower game formulation inspired by the "right-of-way" traffic rules and focuses on the modeling of driver intents and their resulting behaviors under such traffic rules and etiquette. Both approaches lead to interpretable and scalable driver/vehicle interaction models. This dissertation then introduces an application of these models to fast and economical virtual V&V of AV control systems. Secondly, this dissertation presents a high-level control framework for AVs to safely and effectively interact with other road users. The framework is based on a constrained partially observable Markov decision process (POMDP) formulation of the AV control problem, which is then solved using a tailored model predictive control algorithm called POMDP-MPC. The major advantages of this control framework include its abilities to handle interaction uncertainties and provide an explicit probabilistic safety guarantee under such uncertainties. Finally, this dissertation introduces the Action Governor (AG), which is a novel add-on scheme to a nominal control loop for formally enforcing pointwise-in-time state and control constraints. The AG operates based on set-theoretic techniques and online optimization. Theoretical properties and computational approaches of the AG for discrete-time linear systems subject to non-convex exclusion-zone avoidance constraints are established. The use of the AG for enhancing AV safety is illustrated through relevant simulation case studies.