MIT Unveils IntersectionZoo: A New AI Tool for Evaluating Reinforcement Learning in Urban Eco-Driving
Cambridge, MA – In a significant step towards optimizing urban transportation and reducing emissions, researchers at MIT have introduced IntersectionZoo, a novel benchmark system designed to evaluate progress in multi-agent deep reinforcement learning (DRL) for eco-driving. This innovative tool addresses the pressing need for standardized benchmarks in the field, enabling researchers to better assess and improve AI algorithms aimed at enhancing efficiency in complex urban environments.
The motivation behind IntersectionZoo stems from the inefficiencies inherent in city driving, characterized by constant stop-and-go traffic. Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor at MIT, explains, “We got interested a few years ago in the question: Is there something that automated vehicles could do here in terms of mitigating emissions? Is it a drop in the bucket, or is it something to think about?”
Eco-driving, implemented through control systems in autonomous vehicles, offers a promising solution. By making subtle adjustments to minimize unnecessary fuel consumption – such as coasting towards red lights instead of accelerating – these systems can significantly reduce emissions. Moreover, the impact extends beyond individual vehicles, as automated actions influence the behavior of surrounding conventional cars.
IntersectionZoo aims to tackle the challenging optimization problems associated with eco-driving, considering factors like network topology, road grade, temperature, humidity, vehicle types, and fuel mixes. The benchmark comprises 1 million data-driven traffic scenarios, uniquely positioning it to advance DRL generalizability.
A key issue identified by Wu and her team is the lack of generalizability in existing DRL algorithms. Algorithms trained for specific intersections often fail when even minor modifications are introduced, such as adding a bike lane or altering traffic light timing. IntersectionZoo directly addresses this limitation by providing a robust platform for evaluating algorithmic progress under diverse and realistic conditions.
Wu emphasizes that the problem of non-generalizability extends beyond traffic scenarios. “It goes back down all the way to canonical tasks that the community uses to evaluate progress in algorithm design.” By incorporating a rich set of characteristics relevant to real-world problems, IntersectionZoo fills a critical gap in current benchmarking practices.
The research team’s findings were presented at the 2025 International Conference on Learning Representation in Singapore, detailing the design and capabilities of IntersectionZoo.
Looking ahead, Wu and her collaborators plan to apply IntersectionZoo to quantify the impact of eco-driving in automated vehicles on city-wide emissions, considering varying deployment percentages. However, the primary goal of IntersectionZoo is to foster the development of general-purpose DRL algorithms applicable to a wide range of fields, including autonomous driving, video games, security, robotics, warehousing, and classical control problems.
IntersectionZoo and its documentation are freely available on GitHub, empowering researchers worldwide to leverage this valuable tool. The project reflects MIT’s commitment to open research and collaborative innovation.
The IntersectionZoo paper was co-authored by Vindula Jayawardana, Baptiste Freydt, Ao Qu, Cameron Hickert, and Zhongxia Yan.
