IntersectioNDE: Learning Complex Urban Traffic Dynamics based on Interaction Decoupling Strategy

TL;DR

This work addresses the challenge of validating autonomous vehicles in dense urban intersections by introducing the City Crossings Dataset (CiCross) and IntersectioNDE, a data-driven scene-level traffic simulator. Central to the approach is the Interaction Decoupling Strategy (IDS), which trains on disjoint agent subsets to learn marginal interaction dynamics and enables Marginal-to-Joint simulation when generating full-scene futures with a scene-aware Transformer. The method improves robustness and long-term stability in complex, heterogeneous traffic by composing learned interaction primitives, reducing reliance on modeling the full high-dimensional joint distribution. Experiments on CiCross demonstrate improved fidelity to real-world distributions, stronger stability in closed-loop rollouts, and meaningful qualitative cases that reflect realistic urban driving dynamics, suggesting strong potential for scalable AV testing and safety assessment.

Abstract

Realistic traffic simulation is critical for ensuring the safety and reliability of autonomous vehicles (AVs), especially in complex and diverse urban traffic environments. However, existing data-driven simulators face two key challenges: a limited focus on modeling dense, heterogeneous interactions at urban intersections - which are prevalent, crucial, and practically significant in countries like China, featuring diverse agents including motorized vehicles (MVs), non-motorized vehicles (NMVs), and pedestrians - and the inherent difficulty in robustly learning high-dimensional joint distributions for such high-density scenes, often leading to mode collapse and long-term simulation instability. We introduce City Crossings Dataset (CiCross), a large-scale dataset collected from a real-world urban intersection, uniquely capturing dense, heterogeneous multi-agent interactions, particularly with a substantial proportion of MVs, NMVs and pedestrians. Based on this dataset, we propose IntersectioNDE (Intersection Naturalistic Driving Environment), a data-driven simulator tailored for complex urban intersection scenarios. Its core component is the Interaction Decoupling Strategy (IDS), a training paradigm that learns compositional dynamics from agent subsets, enabling the marginal-to-joint simulation. Integrated into a scene-aware Transformer network with specialized training techniques, IDS significantly enhances simulation robustness and long-term stability for modeling heterogeneous interactions. Experiments on CiCross show that IntersectioNDE outperforms baseline methods in simulation fidelity, stability, and its ability to replicate complex, distribution-level urban traffic dynamics.

Figures (6)

• Figure 1: Overview of the training and inference pipeline of IntersectioNDE.
• Figure 2: Architecture of the Scene-Aware Interaction Transformer Model.
• Figure 3: Overview of CiCross map and key traffic statistics.
• Figure 4: Speed Distribution of Different Participants
• Figure 5: Closest Distance Distribution of Different Participants