RLFTSim: Reinforcement Learning Fine-Tuning for Realistic Traffic Simulation

Researchers have developed RLFTSim, a reinforcement-learning-based fine-tuning framework for multi-agent traffic simulation. Unlike supervised open-loop training, which fails to capture dynamic multi-agent interactions, RLFTSim enhances realism by aligning simulator rollouts with real-world data distributions. It also enables goal-conditioned controllability in scenario generation. The framework is instantiated on a pre-trained simulation model with a reward balancing fidelity and controllability. Experiments on the Waymo Open Motion Dataset show state-of-the-art performance in realism. Compared to heuristic search-based methods, RLFTSim requires significantly fewer samples due to a low-variance, dense reward signal.