“`html
If there’s one element that defines driving in any major metropolis, it’s the continual stop-and-start as traffic signals shift and as vehicles, including cars and trucks, converge, diverge, turn, and park. This incessant halting and starting is highly inefficient, elevating the level of pollution, including greenhouse gases, emitted per mile driven.
One method to mitigate this is referred to as eco-driving, which can be implemented as a control system in autonomous vehicles to enhance their efficiency.
What kind of difference could that bring? Would the effects of such systems in lowering emissions justify the investment in the technology? Tackling such inquiries is part of a broad spectrum of optimization challenges that have proven difficult for researchers to confront, and testing their proposed solutions has been complex. These challenges involve many diverse agents, such as various vehicle types in a city, and numerous factors that affect emissions, including velocity, weather, road conditions, and traffic light schedules.
“We became intrigued a few years back by the inquiry: Is there a role that automated vehicles could play in reducing emissions?” remarks Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor in the Department of Civil and Environmental Engineering and the Institute for Data, Systems, and Society (IDSS) at MIT and a principal investigator in the Laboratory for Information and Decision Systems. “Is it merely a drop in the bucket, or is it significant?” she contemplated.
To respond to such a question with multiple components, the first step is to compile all available information about the system from various sources. One source includes the configuration of the network’s topology, Wu explains, which in this case refers to a map illustrating all the intersections within each city. Additionally, there are U.S. Geological Survey data detailing the elevations to ascertain the incline of the roads. Furthermore, temperature and humidity data, along with information on the diversity of vehicle types and ages, as well as fuel type mixtures, are also essential.
Eco-driving entails making minor modifications to decrease unnecessary fuel usage. For instance, as vehicles near a red traffic light, “there’s no reason for me to accelerate towards the red signal,” she states. By merely coasting, “I am not consuming gas or electricity during that time.” If one vehicle, such as an automated one, reduces its speed as it approaches an intersection, the conventional, manual cars trailing behind it will also need to decelerate, thereby extending the benefits of such efficient driving far beyond just the vehicle executing it.
That’s the fundamental concept behind eco-driving, Wu explains. However, to assess the impact of such strategies, “these are complex optimization challenges” that incorporate numerous different factors and parameters, “leading to a surge of interest in how to tackle difficult control problems using AI.”
The novel benchmark system that Wu and her team have developed based on urban eco-driving, designated “IntersectionZoo,” aims to help fulfill part of that need. The benchmark was elaborated in detail in a paper presented at the 2025 International Conference on Learning Representation in Singapore.
Examining methods previously employed to address such intricate problems, Wu notes that a crucial category of approaches is multi-agent deep reinforcement learning (DRL), but a lack of adequate standard benchmarks to assess the outcomes of such methods has hindered advancement within the field.
The new benchmark is designed to tackle a significant issue that Wu and her team recognized two years ago: most existing deep reinforcement learning algorithms, after being trained for a specific scenario (e.g., a particular intersection), fail to maintain relevance when even minor changes are introduced, such as adding a bike lane or altering traffic light timing, even when allowed to train for the modified situation.
Indeed, Wu emphasizes, this issue of non-generalizability “is not confined to traffic,” she comments. “It extends back to fundamental tasks that the community employs to evaluate advancements in algorithm design.” But since most such fundamental tasks do not contemplate modifications, “it becomes challenging to determine if your algorithm is improving regarding this robustness issue if we don’t evaluate for it.”
Although numerous benchmarks are currently utilized to assess algorithmic progress in DRL, she asserts, “this eco-driving dilemma presents a rich array of characteristics that are vital in resolving real-world challenges, especially from the standpoint of generalizability, which no other benchmark fulfills.” This is why the 1 million data-driven traffic scenarios within IntersectionZoo distinctly position it to propel advancements in DRL generalizability. Thus, “this benchmark enhances the diversity of ways to evaluate deep RL algorithms and their progress.”
Regarding the initial question about urban traffic, one focus of ongoing research will be applying this newly crafted benchmarking tool to investigate the specific effects on emissions from implementing eco-driving in automated vehicles within a city, contingent on the percentage of such vehicles that are deployed.
However, Wu emphasizes that “instead of creating a solution that can deploy eco-driving on a citywide scale, the principal objective of this study is to aid in the development of general-purpose deep reinforcement learning algorithms, which can be applied to this application, as well as to all these other applications — autonomous driving, video games, security challenges, robotics, warehousing, classical control issues.”
Wu adds that “the project aims to make this a useful tool for researchers, available openly.” IntersectionZoo, along with documentation on its usage, is accessible at GitHub.
On the paper, Wu is joined by lead authors Vindula Jayawardana, a graduate student in MIT’s Department of Electrical Engineering and Computer Science (EECS); Baptiste Freydt, a graduate student from ETH Zurich; and co-authors Ao Qu, a graduate student in transportation; Cameron Hickert, an IDSS graduate student; and Zhongxia Yan PhD ’24.
“`