Traffic3D is a rich 3D traffic simulation framework designed to bridge the gap between laboratory reinforcement learning research and real-world deployment. Built on a high-fidelity rendering engine, it provides complex multi-intersection scenarios, diverse vehicle behaviours, and a configurable reward structure — enabling researchers to train agents that generalise from simulation to live camera feeds.
Key contributions include a modular agent–environment interface compatible with standard RL frameworks, support for multi-agent scenarios across interacting intersections, and tooling for synthesising training data using Graph Neural Networks to reduce reliance on manual annotation.
Traffic3D has been used as the platform for several PhD research projects at Aston University, and the codebase is publicly available.