MIT’s IntersectionZoo: A New AI Tool to Evaluate Progress in Reinforcement Learning for Eco-Driving
Researchers at MIT have developed a novel benchmark system called “IntersectionZoo” to evaluate progress in multi-agent deep reinforcement learning (DRL), particularly in the context of urban eco-driving. This innovative tool addresses the critical need for standardized benchmarks to assess and improve AI algorithms aimed at optimizing traffic flow and reducing emissions in cities.
Eco-driving, a control system designed for autonomous vehicles, focuses on making subtle adjustments to minimize unnecessary fuel consumption. Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor at MIT, explains that eco-driving involves strategies such as coasting towards red lights instead of accelerating, which can significantly reduce fuel consumption and emissions.
The challenge lies in the complexity of urban traffic systems, which involve numerous interacting agents and factors like vehicle types, weather conditions, road gradients, and traffic light timings. To address this, IntersectionZoo gathers extensive data from various sources, including network topology maps, U.S. Geological Survey data on elevations, and real-time temperature, humidity, vehicle, and fuel type data.
According to the research paper presented at the 2025 International Conference on Learning Representation in Singapore, IntersectionZoo aims to tackle the issue of non-generalizability in existing DRL algorithms. Current algorithms often fail to adapt to even minor changes in traffic scenarios, such as adding a bike lane or adjusting traffic light timings. Wu emphasizes that this problem extends beyond traffic scenarios and affects the broader field of algorithm design.
IntersectionZoo distinguishes itself by offering a rich set of characteristics important for solving real-world problems, particularly from a generalizability standpoint. With over 1 million data-driven traffic scenarios, it uniquely positions itself to advance progress in DRL generalizability, enhancing the ways deep RL algorithms and progress are evaluated.
The tool is openly available for researchers. IntersectionZoo, along with comprehensive documentation, can be accessed on GitHub.
Wu envisions that IntersectionZoo will support the development of general-purpose deep reinforcement learning algorithms applicable not only to eco-driving but also to various other domains, including autonomous driving, video games, security, robotics, warehousing, and classical control problems.
The research team includes 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.
