MIT’s IntersectionZoo: A New Benchmark for Evaluating Progress in Reinforcement Learning

Researchers at MIT have developed a new benchmark system called “IntersectionZoo” to evaluate progress in multi-agent deep reinforcement learning (DRL) for complex urban traffic control problems. This innovative tool aims to address the challenges of optimizing traffic flow and reducing emissions in cities by providing a standardized platform for testing and improving AI algorithms.

The core idea behind IntersectionZoo is to create a realistic simulation environment that captures the intricacies of urban driving, including factors like traffic light timing, vehicle types, road conditions, and weather. Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor at MIT, explains that the system gathers data from various sources, such as U.S. Geological Survey data for road elevations and real-time data on temperature, humidity, and vehicle types.

One of the primary goals of IntersectionZoo is to facilitate the development of eco-driving strategies for autonomous vehicles. Eco-driving involves making small adjustments to driving behavior, such as coasting to red lights instead of accelerating, to minimize fuel consumption and emissions. Wu notes that even small changes in driving behavior by automated vehicles can have a ripple effect, influencing the behavior of non-automated vehicles behind them.

The researchers found that existing DRL algorithms often fail to generalize well to new situations. Even minor modifications to the environment, such as adding a bike lane or changing traffic light timing, can significantly degrade the performance of these algorithms. IntersectionZoo addresses this issue by providing a diverse set of 1 million data-driven traffic scenarios that can be used to train and evaluate DRL algorithms for robustness and generalizability.

The IntersectionZoo benchmark is described in detail in a paper presented at the 2025 International Conference on Learning Representation in Singapore. The tool is freely available on GitHub, allowing researchers worldwide to use it for their own DRL research. Wu emphasizes that the project’s goal is to provide a tool for researchers that is openly available, fostering collaboration and accelerating progress in the field of reinforcement learning.

While the initial focus of IntersectionZoo is on urban eco-driving, the researchers envision that the tool can be applied to a wide range of other applications, including autonomous driving, video games, security problems, robotics, warehousing, and classical control problems. By providing a standardized benchmark for evaluating DRL algorithms, IntersectionZoo has the potential to drive significant advances in AI and its applications across various domains.

Wu is joined on the paper 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.