TU Delft Maps How AI Traffic Simulation Is Quietly Broken
A comprehensive taxonomy maps every AI method being used to simulate how autonomous vehicles and human drivers actually interact on shared roads.
Takeaways
- New survey maps AI methods for simulating mixed AV and human traffic across three method families
- Existing simulators like SUMO and CARLA rely on rule-based models that miss real behavioral variability
- Taxonomy covers single-agent, multi-agent, environment-level, and cognitive/physics-informed approaches
- Core critique: models trained on passive logs cannot handle counterfactual ego interventions
- Open-loop metrics like minADE poorly predict closed-loop simulation quality, causing an evaluation crisis
- Full paper available at arXiv:2604.12857
Testing autonomous vehicles on public roads is dangerous, expensive, and slow. Simulation is the obvious alternative, but a catch has quietly held the field back for years: the traffic surrounding your simulated AV is fake in ways that matter. A new survey from researchers at TU Delft, University of Michigan-Dearborn, and NYU maps the entire landscape of AI-driven approaches trying to fix this, and the picture that emerges is both more advanced and more broken than most practitioners realize.
The survey covers hundreds of methods across three method families, along with benchmarks, datasets, and simulation tools, arguing that the community has been optimizing for the wrong thing.
Why rule-based traffic stopped being enough
Popular simulation platforms like SUMO, VISSIM, and CARLA typically model other vehicles using simple rule-based logic or by replaying recorded trajectories. Classical models such as the Intelligent Driver Model for car-following or MOBIL for lane-changing capture aggregate flow but miss the individual variability and adaptation that defines real driving. That gap widens under mixed autonomy, when AVs are gradually integrated into human-dominated road networks and must navigate behavioral heterogeneity at every intersection.
Previous surveys tackled pieces of the problem. Some cataloged simulators without examining the AI inside them. Others reviewed ego-vehicle decision-making but ignored surrounding traffic. None spanned the full spectrum from single-agent behavior to full scene generation.