PHASE Framework Introduces Self-Play for Realistic Highway Traffic Simulation

A new framework called PHASE (Policy for Heterogeneous Agent Self-play on Expressway) has been developed to generate realistic highway traffic simulations. This context-aware self-play system is designed to address key challenges in autonomous vehicle safety evaluation. It enables broad coverage across various speeds and driving maneuvers while allowing controllable generation of rare, safety-critical scenarios. The framework achieves behavioral credibility in multi-agent interactions through closed-loop multi-agent training. PHASE supports different vehicle profiles, including passenger cars and articulated trailer trucks, within a single policy. This is accomplished via vehicle-aware dynamics and context-conditioned actions. The approach aims to provide scalable safety evaluation for autonomous vehicles by simulating interactions too rare to study from logged data alone. The work is documented in the arXiv preprint 2604.16406v1.