If there is one thing that gives a driving characteristic in any big city, it is constantly stop-and-go while changing traffic lights and the car and truck merges and separated and turn and park. This is constantly closing and starting is extremely inefficient, increasing the amount of pollution, including greenhouse gases, which are emitted per mile of driving.
One approach to dealing with this is known as eco-driving, which can be established as a control system in autonomous vehicles to improve their efficiency.
How much can it make? Would the effect of the systems reduced would be the effect of investing in the technology? Addressing such questions is a wide range of Optim ptimization problems that are difficult for researchers to address, and the solutions they come up with have become difficult to test. These are problems that include many different agents, such as many different types of vehicles and different factors in the city that influence their emissions, including motion, weather, road conditions and traffic light timing.
“We were interested in this question a few years ago: is there something that can be done here in terms of reducing automatic vehicles?” Kathy Wu says, Thomas D and Virginia W. AB Bot Career Development Civil and Environmental Engineering Department Associate Professor and Data at MIT, Systems and Society (IDSS) and Information for Information for Information for Information. “Is it a drop in a bucket, or want to think about?” She was surprised.
To consider such a question associated with many components, the first requirement is to collect all available data about the system from many sources. There is a layout of the topology of a network, Woo says, in this case a map of all intersections in each city. Then to determine the grade of roads, there is a U.S. There is a geological survey data. There is also data on temperature and humidity, the type of vehicle and the combination of the era and on the combination of fuel types.
Eco-driving includes small adjustments to reduce unnecessary fuel consumption. For example, as the car comes closer to the traffic light that has turned red, “I have no sense of driving as fast as possible in red lights,” he says. Only on the coast, “I’m not burning gas or electricity in the meantime.” If a car, such as an automatic vehicle, slows down on an intersection approach, the traditional, non-automatic car behind it will also be forced to slow down, so the effect of such efficient driving can only be expanded outside the car.
That is the original idea behind eco-driving, Woo says. But to find out the effect of such steps, “these are challenging Optim ptimization problems” that include many different factors and parameters, “so how to solve strict control problems using AI is interested right now.”
The new benchmark system that developed on the basis of urban eco-driving, which they call “intersection”, is intended to help keep it in mind. A detailed description of the benchmark was made in a paper presented at a 2025 international conference in Singapore.
Given the approaches used to solve such complex problems, Woo says that an important range methods are multi-agent deep reinforcement learning (DRL), but the lack of sufficient standard benchmarks to evaluate the results of such methods hinders progress in the field.
The new benchmark is intended to consider an important point that WU and its team were identified two years ago, which is that when trained for a particular situation (eg, a specific intersection), when small modifications are added, such as a bike lane or a bike lane for traffic lights.
In fact, Wu points out, this problem of non-normalization is not “specific to traffic,” he says. “It uses the community to evaluate the progress in the algorithm design all the way to the canonical functions.” But because most of these canonical tasks are not involved in changing, “if we do not evaluate it, it is difficult to know if your algorithm is progressing on this type of strength.”
While there are many benchmarks that are currently used to evaluate algorithmic progress in the DRL, it says, “This is a rich set of characteristics in the eco-driving problem that is important in solving real-world problems, especially from the point of view of generalization, and no other benchmark satisfaction.” This is the reason why 1 million data based traffic views in the intersection give it a unique way to further the progress of DRL generalization. As a result, “this benchmark Deep Vand RL algorithms and increases the richness of ways to evaluate progress.”
And when it comes to the initial question about the city’s traffic, one focus of the ongoing work will apply to this newly developed benchmarking tool to consider the special case of applying eco-driving in automatic vehicles in the city, depending on what percentage of such vehicles is deployed.
But Woo added that “instead of creating something that can deploy eco-driving on a city basis, the main goal of this study is to support the development of Deep Perfect Replace Education Algorithms, which can apply to this application, but also all these applications, but also-defeated driving, video games, problems.
Wu adds that “the project is to provide the target as a tool for these researchers, it is openly available.” Documentation on intersection, and how to use it is freely available at Github.
Wu paper is joined by lead writers Vindula Jayawardhan, a graduate student of MIT’s Electrical Engineering and Computer Science (EEC) department; Baptist Fredt, Ath Zurich’s graduate student; And co-authors AO Q, a graduate student of transport; Cameron Hikirt, IDSS Graduate Student; And Zhongxia Yan Ph.D.24.
