Cooperative Multi-Agent Sensing, Planning, and Control for Connected Autonomous Vehicles

Abstract

Modern transportation networks are witnessing ever-increasing levels of vehicle automation as subsystems originally controlled by the human driver are taken over by automated control systems. Connected autonomous vehicles (CAVs) represent the latest technological breakthrough in this domain, which promises increased safety, improved efficiency, and better accessibility. Recently, more and more academic and industrial research interest has been focused on guaranteeing the safe operation of a single CAV by harnessing its advanced sensing and computation capability. However, CAVs also have another very important capability whose potential remains largely untapped: communication. Communication among CAVs (V2V - Vehicle to Vehicle) and with infrastructure (V2I - Vehicle to Infrastructure) provide CAVs the ability to share progressively gathered information, policies, and rules, and pursue controls related to both local and global objectives (in space and time). Thus, CAVs can collaboratively achieve levels of safety and efficiency that cannot be achieved by a single vehicle working alone. With this new focus on cooperative CAV control, we propose methods and associated algorithmic implementations, through which CAVs can contribute to solving some of the persistent problems which have been plaguing road and highway infrastructure for decades. We provide problems such as highway merge junction bottlenecks, highway traffic shock waves, and signalized or unsignalized intersection management with vastly superior solutions by harnessing the communication capabilities of CAVs. In both the domains of fully autonomous traffic and mixed traffic (CAVs coexisting with human driven vehicles), communication-based CAV control allows for safer operation with improved global throughput of traffic flow. In a world facing increasing shortages of traditional fuel sources, the ability of cooperative CAV control to reduce overall fuel consumption levels is also quite notable. The algorithms proposed in this dissertation, which use heuristic, optimization, and learning-based control, can be broadly categorized as cooperative multi-agent control in which CAV sensing and/or actuation data are communicated. Sharing of this data allows CAVs to gather information about downstream traffic conditions, and overcome problems such as hidden obstacles due to sensor occlusion and dangerous conditions ahead or out of a single vehicle’s “line of sight”. This increased collection of information available to the control center (either ego vehicle or infrastructure) can then be used to compute better controls leading to safer, more efficient operation, and the ability to prevent the occurrence of certain unwanted traffic conditions. This potential capability of CAVs is explored using both centralized (often V2I-based) and decentralized (often V2V-based) control strategies. Moreover, as ensuring the safe operation of the overall system is a primary concern, in addition to achieving the desired performance, safety is a fundamental constraint built into all these methods. Our overall goal is to provide practically implementable algorithms that achieve high levels of performance and are capable of real-time operation while demonstrating robustness to variations in the environment. As such, we pay special attention to realistic driving environments including the impact of features such as multiple lanes, mixed traffic, heterogeneous traffic (cars, buses, trucks, emergency vehicles, etc.), and limitations in both communication and computation resources. We investigate the properties of communication networks used, as well as the ability of the control systems to adapt to imperfections in actuation, sensing, and communication.