Traffic Control via Connected and Automated Vehicles: An Open-Road Field Experiment with 100 CAVs

TL;DR

This work tackles stop-and-go traffic instabilities caused by human driving by introducing MegaController, a hierarchical, modular control framework that couples a centralized Speed Planner with decentralized Vehicle Controllers to smooth traffic in real road conditions. The Speed Planner fuses INRIX data with AV-ping observations, enhances traffic state estimation, and designs lane-specific speed targets using kernel smoothing, MPC, and reinforcement learning, while Vehicle Controllers implement either acceleration-based or ACC-based control with lane-change safety safeguards. The authors develop both macroscopic mean-field traffic models and finite/infinite-dimensional optimal control formulations, validate controllers in simulations, and perform the largest open-road field test with 100 CAVs on I-24 MOTION (MegaVanderTest), demonstrating measurable reductions in stop-and-go waves and energy inefficiency. They also detail a comprehensive data pipeline, energy modeling framework, and a public data release to enable replication and further study, underscoring the practical viability and societal potential of CAVs for traffic efficiency.

Abstract

The CIRCLES project aims to reduce instabilities in traffic flow, which are naturally occurring phenomena due to human driving behavior. These "phantom jams" or "stop-and-go waves,"are a significant source of wasted energy. Toward this goal, the CIRCLES project designed a control system referred to as the MegaController by the CIRCLES team, that could be deployed in real traffic. Our field experiment leveraged a heterogeneous fleet of 100 longitudinally-controlled vehicles as Lagrangian traffic actuators, each of which ran a controller with the architecture described in this paper. The MegaController is a hierarchical control architecture, which consists of two main layers. The upper layer is called Speed Planner, and is a centralized optimal control algorithm. It assigns speed targets to the vehicles, conveyed through the LTE cellular network. The lower layer is a control layer, running on each vehicle. It performs local actuation by overriding the stock adaptive cruise controller, using the stock on-board sensors. The Speed Planner ingests live data feeds provided by third parties, as well as data from our own control vehicles, and uses both to perform the speed assignment. The architecture of the speed planner allows for modular use of standard control techniques, such as optimal control, model predictive control, kernel methods and others, including Deep RL, model predictive control and explicit controllers. Depending on the vehicle architecture, all onboard sensing data can be accessed by the local controllers, or only some. Control inputs vary across different automakers, with inputs ranging from torque or acceleration requests for some cars, and electronic selection of ACC set points in others. The proposed architecture allows for the combination of all possible settings proposed above. Most configurations were tested throughout the ramp up to the MegaVandertest.

Figures (13)

• Figure 1: Architectural framework of the MegaController. The hierarchical and modular nature of the design allows for greater flexibility in design decisions and dealing with varied sensing and actuation capabilities of the heterogeneous fleet. The blue box represents the centralized Speed Planner unit, and the red boxes represent decentralized Vehicle Controllers, which are vehicle-dependent (that is, each vehicle has a different control architecture and thus requires a different control paradigm). The components work in concert to achieve higher level goals of flow smoothing.
• Figure 2: The Markov Decision Process that RL is based on. An agent exists in an environment and repeatedly chooses action based upon a state and receives rewards, which then informs the agent on the value of the state and action pair.
• Figure 3: Data Pipeline and Major Function Modules of Speed Planner At the beginning of each update, the Speed Planner extracts a combination of macroscopic TSE and vehicle observations from the corresponding factual tables (fact_inrix_estimate, fact_vehicle_ping) in the database to calculate the target speed profile. The raw TSE is used as the input of the prediction module, of which the output is fused with vehicle observations. The fusion is then smoothed and used in the buffer design module, of which the output is saved into the database (fact_speed_planner) and published as the target speed profile.
• Figure 4: An example of the effect of the lane-change handling mechanism on a real-world trajectory in the event of a cut-in. Top left: lane-change event is detected and considered safe so the lane-change recovery controller was active for 2.9s. Top right: a car cuts-in in front of the ego vehicle at a headway of 65m. Bottom left: the main controller commanded acceleration drops sharply due to the lane-change event causing a large jerk value, but the lane-change controller smooths out this drop in the acceleration. Bottom right: The relative velocity is large enough at the lane-change allowing the controller to be active.
• Figure 5: A photograph of a Nissan Rogue's steering wheel buttons that control the vehicle's ACC. The ACC system is turned on by manually pressing the blue icon on the far right. Through our vehicle interfacing efforts, we are able to electronically press the $+$ and $-$ buttons to toggle the ACC speed setting up 1 mph and down 1 mph, respectively, or hold them to increment by 5 mph. The three ACC gap settings are rotated through by pressing the button on the bottom right.