Traffic congestion is rarely caused by a lack of road space; it is often a byproduct of human driving behavior. Small perturbations—a sudden tap on the brakes or an aggressive lane change—ripple through a line of cars, amplifying into "phantom traffic jams" and fuel-wasting stop-and-go waves. Traditionally, we have tried to solve this with static infrastructure like traffic lights and ramp meters.
A growing body of research suggests a more dynamic alternative: using a small number of Connected and Automated Vehicles (CAVs) as "mobile actuators" to subtly choreograph the behavior of the human drivers around them. By controlling the pace of the few, we can regulate the flow of the many.
The Autonomous Vehicle as a Longitudinal Filter
The most fundamental way an autonomous vehicle influences traffic is through speed control. In a typical traffic stream, human drivers react to the car in front of them with a slight delay and often over-correct, leading to instability. Yue Zhang describes CAVs as "longitudinal filters" that dampen these erratic speed changes.
By maintaining a smooth, calculated velocity, a single CAV can prevent the formation of shockwaves. This doesn't just benefit the autonomous car; Liu et al. demonstrate that this creates "second-hand" fuel savings for the human-driven vehicles following in its wake. When the lead vehicle moves at a stable pace, the followers are forced to do the same, reducing the energy-intensive cycle of braking and accelerating. Laura Yu Zheng and Ming Lin frame this as "Traffic-Aware Autonomous Driving," where the vehicle acts as a "pace car" to optimize global goals like total traffic velocity rather than just its own travel time.
Mobile Infrastructure at Intersections
The "mobile actuator" concept becomes even more powerful at intersections. Typically, intersections require expensive infrastructure like traffic signals to manage conflicting flows. However, Jia Pan and Weizi Li have shown that "Robot Vehicles" (RVs) can function as mobile, virtual traffic signals.
By strategically stopping or slowing down at an unsignalized junction, an RV can block a stream of human drivers to allow cross-traffic to pass. This decentralized control can reduce emissions by up to 43% by eliminating the rigid, inefficient cycles of traditional lights. Similarly, Shengchao Yan and Wolfram Burgard treat autonomous agents as "courteous" mobile infrastructure that yields or slows down to coordinate traffic patterns, ensuring both efficiency and fairness for human drivers who have no idea they are being "managed."
Choreography through Lateral Control
While longitudinal control manages the speed of a lane, lateral control—the ability to change lanes—allows AVs to manage the capacity of the road. Jia Pan and Weizi Li expanded their framework to include "choreography," where RVs use lane changes to dynamically block or open lanes.
Imagine a highway merge: an AV can move into a specific lane to create a gap for merging traffic or slow down a specific lane to balance the density across the entire road. Chiara Daini and Antonella Ferrara describe these vehicles as "moving bottlenecks." By intentionally imposing local capacity constraints, a tiny fleet of just five vehicles can regulate the bulk flow of an entire network, smoothing out the "stop-and-go" waves that characterize heavy congestion.
Scaling from Simulation to the Real World
The challenge of this research is that human behavior is notoriously difficult to model mathematically. To bypass this, researchers like Jiawei Wang and Yuning Jiang have developed data-driven predictive control strategies that don't require a perfect model of human drivers. Instead, they use local, real-time data to smooth traffic.
To prove these theories work outside of a computer, Kathy Jang, Alexandre M Bayen, and Andreas A. Malikopoulos demonstrated "zero-shot" policy transfer. They trained a reinforcement learning agent in a simulator and placed it directly into a physical 1:25 scale model city. The autonomous vehicle successfully acted as a platoon leader at a roundabout, reducing travel times for the human-driven scale cars by 22% without any prior "real-world" training.
This suggests that we don't need a 100% autonomous fleet to see the benefits of self-driving technology. By treating the first generation of AVs as active participants in traffic management, we can transform them from mere passengers into the "sheepdogs" of the highway.