MIT’s IntersectionZoo: A New Benchmark for Evaluating Reinforcement Learning in Urban Eco-Driving
Researchers at MIT have unveiled IntersectionZoo, a novel benchmark system designed to evaluate the progress of multi-agent deep reinforcement learning (DRL) algorithms in addressing complex urban eco-driving challenges. This innovative tool aims to optimize traffic flow and reduce emissions by making AI-driven adjustments to vehicle behavior at intersections.
The core concept behind eco-driving involves making subtle modifications to driving patterns to minimize unnecessary fuel consumption. Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor at MIT, explains, “there’s no point in me driving as fast as possible to the red light. By just coasting, I am not burning gas or electricity in the meantime.” This approach not only benefits individual vehicles but also influences the behavior of surrounding cars, amplifying the overall impact.
IntersectionZoo addresses a critical gap in the field of DRL by providing a standardized platform for evaluating the generalizability of algorithms. Wu and her team discovered that existing DRL algorithms often fail to adapt when even minor changes are introduced to the environment, such as adding a bike lane or adjusting traffic light timing. This lack of robustness hinders the practical application of these algorithms in real-world scenarios.
The benchmark comprises 1 million data-driven traffic scenarios, offering a rich and diverse set of conditions for training and testing DRL algorithms. This extensive dataset enables researchers to assess the performance of algorithms across a wide range of situations and identify areas for improvement.
According to the research paper presented at the 2025 International Conference on Learning Representation in Singapore, IntersectionZoo uniquely positions itself to advance progress in DRL generalizability. By focusing on the eco-driving problem, which encompasses various real-world complexities, the benchmark offers a more comprehensive evaluation of algorithmic performance.
The ultimate goal of IntersectionZoo is to foster the development of general-purpose DRL algorithms that can be applied to a wide array of applications beyond eco-driving, including autonomous driving, video games, security, robotics, warehousing, and classical control problems.
IntersectionZoo and its associated documentation are freely available on GitHub, providing researchers with a valuable resource for advancing the field of reinforcement learning.
The research team includes lead authors Vindula Jayawardana, Baptiste Freydt, and co-authors Ao Qu, Cameron Hickert, and Zhongxia Yan.
