UniMM: A Unified Mixture Model Framework for Multi-Agent Simulation

The UniMM framework introduces a unified mixture model for generating multimodal behaviors in multi-agent simulations, addressing challenges like behavioral multimodality and distributional shifts. It incorporates a closed-loop sample generation method and a temporal disentanglement-and-alignment mechanism to enhance realism and mitigate learning issues. The framework's distinct model variants, including discrete and anchor-based models, achieve state-of-the-art results on the WOSAC benchmark, offering practitioners improved tools for simulating autonomous driving systems.