A Generative Model for Closed-Loop Microsimulation of Signalized Intersections

Researchers present Enactor, a generative AI model designed to simulate vehicle behavior at signalized intersections with improved accuracy over existing methods. The model uses transformer-based architecture to predict vehicle trajectories in closed-loop simulations, achieving significantly better performance on safety metrics and traffic flow distribution compared to baseline approaches.

Enactor addresses a critical limitation in traffic microsimulation: existing hand-crafted behavior models fail to capture heterogeneous vehicle interactions, while learned predictors suffer from instability when run continuously. This research bridges that gap by developing an actor-centric generative model that maintains stability across extended 4000-second simulations while accurately reproducing real traffic patterns. The model's use of polar coordinates centered on the intersection and separate spatial-temporal attention mechanisms enables it to understand both immediate vehicle dynamics and broader traffic flow patterns.

The advancement matters because accurate traffic simulation directly impacts urban planning, autonomous vehicle development, and intersection safety optimization. Current industry approaches rely on either oversimplified physics-based models or machine learning systems that fail under realistic, closed-loop conditions. Enactor's demonstrated improvements—achieving KL divergence roughly 5-10 times lower than transformer baselines on key metrics—represent substantial progress toward deployable simulation tools that developers and city planners can rely on.

The research has implications for autonomous vehicle developers who use simulations to validate safety systems, and for traffic engineers designing intersection management strategies. Validation against both synthetic SUMO data and real-world fish-eye camera footage demonstrates the model's generalizability. The identified importance of the leader rear-bumper feature as a critical safety-awareness component provides actionable insights for model design in related applications.

Future development should focus on scaling Enactor to multi-intersection networks and incorporating weather/lighting variations present in real deployment environments. The model's performance on real-world field data validation suggests practical feasibility, though edge cases in heavy congestion scenarios warrant further investigation.

• →Enactor achieves 5-10x lower KL divergence compared to transformer baselines on traffic flow metrics
• →The model successfully maintains stability across extended 4000-second closed-loop simulations at multiple intersection geometries
• →Validated on both synthetic SUMO data and real-world traffic footage from fish-eye cameras
• →Reduces red-light violations by more than an order of magnitude relative to baseline approaches
• →Leader rear-bumper distance emerged as the single most impactful feature for safety-aware vehicle behavior