Drivers who have endured several red-light cycles know the frustration of waiting at busy intersections. Beyond testing patience, this idle time may account for up to 15% of carbon dioxide emissions from U.S. land transportation.
An extensive modeling study by MIT researchers shows that eco-driving strategies—such as adjusting speeds to minimize stops and sharp accelerations—could substantially cut these emissions.
By applying a sophisticated AI approach known as deep reinforcement learning, the team analyzed key factors influencing vehicle emissions across three major U.S. cities.
Eco-Driving Could Significantly Cut Intersection Emissions Without Disrupting Traffic
The study suggests that fully implementing eco-driving strategies could lower annual carbon emissions from city intersections by 11–22% without reducing traffic flow or compromising road safety.
The researchers also found that if just 10% of vehicles adopt eco-driving, it could achieve 25–50% of the total potential CO₂ reduction.
Furthermore, adjusting speed limits dynamically at roughly 20% of intersections could deliver 70% of the overall emission benefits. This shows that eco-driving can be introduced gradually while still producing meaningful gains for climate change mitigation and public health.
Image Credits: Courtesy of the researchers
“Approaches such as eco-driving can play a significant role in reducing climate change impacts,” says senior author Cathy Wu, Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS) at MIT, and a member of the Laboratory for Information and Decision Systems (LIDS). “Our work demonstrates that advanced machine-learning techniques, like deep reinforcement learning, can speed up the analyses that inform sociotechnical decisions — and this is only scratching the surface.”
Her co-authors include lead author Vindula Jayawardana, an MIT graduate student, along with MIT graduate students Ao Qu, Cameron Hickert, and Edgar Sanchez; MIT undergraduate Catherine Tang; Baptiste Freydt, a graduate student at ETH Zurich; and Mark Taylor and Blaine Leonard from the Utah Department of Transportation. The study was published in Transportation Research Part C: Emerging Technologies.
A Comprehensive, Multi-Phase Modeling Study
When people think of traffic control, they often picture fixed infrastructure such as stop signs and traffic lights. However, advancements in vehicle technology are opening the door to eco-driving — a broad term for vehicle-based traffic management strategies, like adjusting speeds dynamically to reduce energy use.
In the short term, eco-driving could take the form of speed recommendations delivered via vehicle dashboards or smartphone apps. In the future, it could evolve into intelligent speed controls that directly manage the acceleration of semi-autonomous and fully autonomous vehicles through vehicle-to-infrastructure communication systems.
“Much of the earlier research has centered on how to implement eco-driving,” Wu explains. “We reframed the question to ask whether we should implement it. If this technology were deployed on a large scale, would it truly make an impact?”
Comprehensive Four-Year Study Identifies 33 Key Factors Influencing Vehicle Emissions
To explore this, the team undertook an extensive, multi-year modeling study that spanned nearly four years.
Their first step was to pinpoint 33 factors affecting vehicle emissions, such as temperature, road grade, intersection layout, vehicle age, traffic volume, vehicle types, driver behavior, signal timing, and road geometry, among others.
“One of the greatest challenges was ensuring we were thorough and didn’t overlook any key factors,” Wu says.
The team then drew on data from OpenStreetMap, U.S. geological surveys, and other sources to build digital models of over 6,000 signalized intersections across three cities — Atlanta, San Francisco, and Los Angeles — and simulated more than a million traffic scenarios.
Using deep reinforcement learning, they optimized each scenario for eco-driving to maximize emissions reductions.
Reinforcement learning fine-tunes vehicle driving behavior through trial-and-error within a high-fidelity traffic simulator, rewarding energy-efficient actions and penalizing inefficient ones.
Decentralized Cooperative Approach Tackles Diverse Intersection Traffic Scenarios
The team framed the challenge as a decentralized, cooperative multi-agent control problem, in which vehicles work together to improve overall energy efficiency — even benefiting non-participating vehicles — while operating independently, eliminating the need for costly vehicle-to-vehicle communication.
A key difficulty was training behaviors that could generalize across varied intersection traffic scenarios. The researchers noted that certain scenarios shared similarities, such as having the same number of lanes or identical traffic signal phase patterns.
To address this, the researchers trained separate reinforcement learning models for distinct clusters of traffic scenarios, which led to greater overall emissions reductions.
Even with AI assistance, Wu notes, analyzing an entire city’s traffic network at once would be so computationally demanding that it might take another decade to complete.
Instead, they broke the challenge into smaller parts, tackling each eco-driving scenario at the level of individual intersections.
“We placed strict limits on how eco-driving controls at one intersection could affect nearby intersections,” Wu explains. “This approach greatly simplified the problem, allowing us to conduct the analysis at scale without introducing unpredictable network effects.”
Substantial Emissions Reductions
The analysis revealed that fully implementing eco-driving could cut intersection-related emissions by 11–22%.
The scale of these benefits varies with a city’s street layout. For example, in dense cities like San Francisco, the limited distance between intersections leaves less opportunity for eco-driving, potentially reducing savings. In contrast, Atlanta’s higher speed limits could yield greater reductions.
Even with only 10% of vehicles practicing eco-driving, cities could achieve 25–50% of the maximum emissions benefit thanks to car-following effects, where non-eco-driving vehicles mirror the smoother, optimized speeds of eco-driving vehicles, lowering their own emissions in the process.
In some instances, eco-driving could boost vehicle throughput while reducing emissions. However, Wu warns that higher throughput might encourage more drivers to use the roads, potentially diminishing the emissions gains.
Their evaluation of common safety indicators — known as surrogate safety measures, such as time to collision — suggests that eco-driving is as safe as human driving, though it may lead to unexpected reactions from human drivers. Wu notes that further research is needed to fully assess potential safety implications.
Combining Eco-Driving with Other Green Technologies Yields Greater Emission Reductions
The findings also indicate that eco-driving can deliver even greater results when paired with other transportation decarbonization strategies. For example, in San Francisco, 20% adoption of eco-driving alone could cut emissions by 7%, but when combined with projected hybrid and electric vehicle adoption, the reduction could reach 17%.
“This marks the first systematic effort to quantify the environmental benefits of eco-driving across an entire network,” says Hesham Rakha, the Samuel L. Pritchard Professor of Engineering at Virginia Tech, who was not involved in the study. “It’s an important piece of research that will serve as a key reference for future assessments of eco-driving systems.”
Although the team’s analysis centers on carbon emissions, the gains closely align with reductions in fuel consumption, energy use, and improvements in air quality.
“Eco-driving is almost a cost-free solution,” Wu adds. “Most cars already have smartphones on board, and vehicles with advanced automation features are becoming more common. For a solution to scale quickly in the real world, it needs to be simple to deploy and ready to go — and eco-driving meets that standard.”
Read the original article on: MIT
Read more: Self-driving Maserati breaks autonomous speed record