Safety-critical multi-agent control and coordination for autonomous vehicles: barrier functions, learning, and deployment
OA Version
Citation
Abstract
The emergence of Connected and Automated Vehicles (CAVs), coupled with advancements in traffic infrastructure, promises to address longstanding transportation challenges, including accidents, congestion, energy inefficiency, and environmental pollution. Achieving this vision critically depends on effective traffic management, particularly at bottlenecks such as intersections, roundabouts, and merging roadways. This dissertation advances safety-critical multi-agent control and coordination for autonomous vehicles through a sequence of integrated contributions. First, motivated by the limitations of conventional time-driven control methods, event-triggered and self-triggered control strategies are developed for CAVs using Control Barrier Functions (CBFs) to enforce hard safety constraints in traffic conflict zones. These methods address infeasibility in solving Quadratic Programs (QPs) by adaptively adjusting control updates based on well-defined and observable system events. Second, motivated by the lack of accurate human behavior models in mixed traffic, the problem of merging an autonomous vehicle (AV) into flows of Human-Driven Vehicles (HDVs) is addressed. An optimal index policy is derived to minimize travel time and energy, and the approach is extended to arbitrary CAV penetration rates through a hierarchical framework that selects safe merging sequences without requiring human modeling, followed by decentralized motion planning under safety constraints. Third, motivated by the challenge of manual parameter tuning in safety-critical controllers, a bilevel Reinforcement Learning (RL) framework is proposed to optimize parameters within an MPC-CBF architecture. The RL agent learns cost weights and CBF constraint parameters without backpropagating through the optimization layer, improving feasibility and performance. The learned controllers are validated in a multi-agent merging scenario and deployed in a Hardware-in-the-Loop (HIL) Smart City Testbed integrating both physical robots and simulated vehicles. Finally, motivated by the need for scalable and decentralized coordination in complex traffic systems, HMARL-CBF is introduced—a Hierarchical Multi-Agent Reinforcement Learning framework that selects high-level cooperative skills and ensures pointwise safety through low-level CBF-based control. This approach demonstrates strong safety compliance and cooperative performance in dense multi-agent environments, outperforming state-of-the-art learning baselines. Together, this dissertation establishes a scalable and adaptive framework for deploying CAV technologies in complex, mixed-autonomy settings. By addressing feasibility limitations in CBF-based controllers, optimizing control parameters through learning-based methods, and validating robustness through hardware-in-the-loop deployment, it enables both theoretical rigor and practical effectiveness, paving the way toward safer and more efficient transportation systems.
Description
2025