If there is one thing that characterizes driving in a big city, it is constant stop as the traffic lights change and the cars and trucks merge and separate and turn. This constant stop and start -ups are extremely ineffective, which increases the amount of pollution, including greenhouse gases, which is issued by driving mile.
An approach to counter this is known as eco-driving, which can be installed as a control system in autonomous vehicles to improve their efficiency.
What difference could it do? Was the impact of these systems in the reduction of emissions worth the investment in technology? Tackling such questions is part of a large category of optimization problems that have been difficult to solve for researchers, and it was difficult to test the solutions they offer. These are problems that involve many different agents, such as the many types of vehicles in a city, and various factors that influence their emissions, including speed, weather, road conditions and traffic shirts.
“We were interested in the question a few years ago: is there something that automated vehicles could do here in terms of attenuation of emissions?” Said Cathy Wu, Associate Professor of Thomas D. and Virginia W. Cabot in career development in the Department of Civil and Environmental Engineering and the Institute of Data, Systems and Society (IDSS) at MIT, and a principal researcher in the information laboratory and decision -making systems. “Is it a drop in the bucket, or is it something to think?” She wondered.
To answer such a question involving as many components, the first requirement is to collect all the data available on the system, from many sources. One is the provision of the network topology, known as Wu, in this case a map of all the intersections of each city. Then there are data on the American geological raising showing the elevations, to determine the note of the roads. There are also data on temperature and humidity, data on the mixture of vehicles and ages, and on the fuel type mixture.
The eco-driving involves making small adjustments to minimize unnecessary fuel consumption. For example, as cars approach a traffic light that has become red, “I would not be by driving as quickly as possible towards the red light,” she said. In the event of a journey, “I do not burn gas or electricity in the meantime.” If a car, like an automated vehicle, slows down as an intersection approaches, then conventional cars not automated behind them will also be forced to slow down, so that the impact of such effective driving can extend far beyond the car that does it.
This is the basic idea behind eco-driving, said WU. But to determine the impact of these measures, “these are difficult optimization problems” involving many different factors and parameters, “there is therefore a wave of interest at the moment to solve hard control problems using AI”.
The new reference system that WU and its collaborators have developed on the basis of urban eco-driving, which they call “Intersectionzoo”, aims to respond to part of this need. The reference has been described in detail in a paper Presented at the 2025 international conference on the representation of learning in Singapore.
By examining the approaches that have been used to solve these complex problems, WU claims that an important category of methods is learning in multi-agent deep strengthening (DRL), but a lack of adequate standard references to assess the results of these methods has hampered progress in the field.
The new reference is intended to solve an important problem that WU and its team identified two years ago, that is to say with most of the apprenticeship algorithms in existing deep reinforcement, when they are trained for a specific situation (for example, a particular intersection), the result does not remain relevant when even small modifications are made, such as bicycle training or the modification of the moment of traffic, even when they are authorized to Train for modified training or to modify the moment of traffic light, even when authorized to train for modified training.
In fact, WU points out that this problem of non-generalization “is not unique to traffic”, she says. “This gives back to the canonical tasks that the community uses to assess progress in the design of algorithms.” But because most of these canonical tasks do not imply making changes, “it is difficult to know if your algorithm progresses on this type of robustness problem, if we do not assess that.”
Although there are many benchmarks that are currently used to assess algorithmic progress in DRL, it says: “This eco-communator problem presents a rich set of characteristics which are important to solve the problems of the real world, in particular from the point of view of generalization, and that no other benchmark satisfies.” This is why the 1 million traffic scenarios based on data in Intersectionzoo position it uniquely to advance the progress of the generalization of DRL. Consequently, “this reference adds to the richness of the means of assessing deep algorithms and progress RL.”
And with regard to the initial question on municipal traffic, an in progress work objective will be to apply this newly developed comparative analysis tool to meet the particular case of the emissions would come from the implementation of eco-driving in automated vehicles in a city, depending on the percentage of these vehicles which are really deployed.
But Wu adds that “rather than doing something that can deploy eco -driving on a city scale, the main objective of this study is to support the development of learning algorithms in deep reinforcement for general use, which can be applied to this application, but also to all these other applications – autonomous driving, video games, security problems, robotics problems, warehouse, classic control problems.”
Wu adds that “the project objective is to provide this as a tool for researchers, which is openly available”. Intersectionzoo, and the documentation on how to use it, are available for free Github.
Wu is joined on the article by the main authors Vindula JayaWardana, a student graduated from the MIT electrical and computer engineering department (EECS); Baptiste Freydt, a student graduated from Eth Zurich; and the co-authors AO, a student graduated by transport; Cameron Hickert, a student graduated from IDSS; and Zhongxia Yan Phd '24.
