This project assesses the performance benefits of implementing multi-agent reinforcement learning (MARL)-based cooperative platooning and cooperative adaptive cruise control with unconnected vehicles (CACCu) in mixed traffic environments, comparing scenarios with and without these technologies under various connected automated vehicle (CAV) market penetrations. The main goal is to investigate when a policy to deploy these advanced technologies makes sense.
Cooperative platooning in mixed traffic, where CAVs must interact safely and efficiently with human-driven vehicles, remains a key barrier to realizing the full mobility, safety, and energy benefits of connected automation, a challenge amplified by uncertainty in human driving behavior. When a CAV’s immediate preceding vehicle is not connected, it may benefit from a lane change to follow a connected vehicle and form cooperative adaptive cruise control; the team’s MARL approach, built on a CNN QMIX architecture supporting centralized training with decentralized execution, learns coordination policies that adapt to surrounding vehicles rather than relying on fixed rules.
The project will deliver a literature review of CAV cooperative platooning technologies in mixed traffic; the design and evaluation of new MARL algorithms that enable dynamic, secure cooperative platooning among CAVs in mixed traffic, building on the team’s CNN-based QMIX work and extending it toward graph-neural-network-based MARL to handle network geometry changes such as lane drops, merges, diverges, and weaving sections; and an evaluation using the open-source SUMO simulator, into which a vehicle dynamics model will be integrated to ensure realistic mixed-traffic behavior. Human-driven vehicles will be modeled at three levels of driving aggressiveness, and various CAV and connected vehicle market penetrations will be considered to quantify the benefits of each technology individually and combined. Outputs include validated performance metrics across traffic flow efficiency, safety, energy consumption, and travel time, delivered as open-source software packages with documentation.
Universities Involved
University of Virginia
Principal Investigators
B. Brian Park
Expected Research Outcomes & Impacts
This research will deliver quantifiable evidence on when cooperative platooning and CACCu technologies become operationally beneficial, establishing simulation-based thresholds for CAV market penetration that justify deployment. Expected outcomes include validated performance metrics across traffic flow efficiency, safety (crash-risk reduction and string stability), energy consumption, and travel time under realistic mixed traffic conditions with varying human driving aggressiveness, supporting evidence-based deployment policy for these advanced technologies.
Technology transfer will occur through open-source software packages with documentation, enabling state DOTs, municipalities, and private-sector partners to evaluate and adapt the MARL-based cooperative platooning algorithms and CACCu control frameworks. Direct engagement with transportation agencies through the Commonwealth Cyber Initiative network will include demonstration workshops and technical briefings showcasing the simulation-based assessment methodology, while the project also supports workforce development by training graduate students in multi-agent reinforcement learning, cooperative vehicle control, and mixed-traffic simulation.
Subject Areas
Connected and Automated Vehicles, Reinforcement Learning, Vehicle Platooning, Traffic Simulation, Mixed Traffic Operations
