IntersectionZoo: New AI Benchmark Tool Assesses Progress in Reinforcement Learning for Eco-Driving
Researchers at MIT have developed a new benchmark system called “IntersectionZoo” to evaluate progress in multi-agent deep reinforcement learning (DRL) for urban eco-driving. This tool aims to address the challenges of optimizing complex systems involving multiple agents, such as vehicles in a city, and various factors influencing their emissions.
Eco-driving, a control system for autonomous vehicles, seeks to minimize unnecessary fuel consumption by making small adjustments like coasting towards red lights instead of accelerating. The impact of such systems on reducing emissions has been difficult to quantify due to the complexity of urban environments.
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?”
IntersectionZoo addresses the need for standardized benchmarks in DRL. It provides 1 million data-driven traffic scenarios to evaluate algorithms and their generalizability. The tool is designed to test how well algorithms adapt to modifications in the environment, such as adding a bike lane or changing traffic light timing.
According to the research paper, existing DRL algorithms often fail to remain relevant when even small modifications are made to the training environment. This lack of generalizability is a significant issue in the field.
IntersectionZoo uniquely positions itself by offering a rich set of characteristics important in solving real-world problems, particularly from a generalizability point of view. The benchmark includes data on network topology, road grades, temperature, humidity, vehicle types, and fuel types.
The tool is intended to support the development of general-purpose deep reinforcement learning algorithms applicable to various fields, including autonomous driving, video games, security problems, robotics, warehousing, and classical control problems.
IntersectionZoo, along with documentation, is freely available on GitHub.
The research was presented at the 2025 International Conference on Learning Representation in Singapore and involves contributions from Vindula Jayawardana, Baptiste Freydt, Ao Qu, Cameron Hickert, and Zhongxia Yan.
</n
