Revisit Mixture Models for Multi-Agent Simulation: Experimental Study within a Unified Framework

Longzhong Lin1,2, Xuewu Lin2, Kechun Xu1, Haojian Lu1, Lichao Huang2, Rong Xiong1, Yue Wang1
1Zhejiang University, 2Horizon Robotics
Paper Code

Multi-Agent Simulation: Example Scenarios in the Waymo Open Sim Agents Challenge (WOSAC)

Overview

MY ALT TEXT

For the main challenges in multi-agent simulation, specifically multimodality and distributional shifts, we formulate research questions centered on mixture models and closed-loop sample generation. To investigate the hypothesis and questions, we revisit the unified mixture model (UniMM) framework, recognizing critical configurations from both model and data perspectives. Building on our exploration, the distinct variants proposed within the UniMM framework, including discrete, anchor-free, and anchor-based models, all achieve state-of-the-art performance.

Quantitative Results

Qualitative Results

Note: Open/Closed-Loop indicates that the model is trained using open-loop or closed-loop samples. K denotes the number of components. Tpred denotes the prediction horizon. Tz* denotes the positive matching horizon. Tpost denotes the posterior planning horizon. τ denotes the simulation update interval.

• Anchor-Free, Open-Loop, K = 6, Tpred = 0.5s

• Anchor-Free, Open-Loop, K = 6, Tpred = 4s

• Anchor-Free, Open-Loop, K = 16, Tpred = 4s

• Anchor-Free, Closed-Loop, K = 6, Tpred = Tz* = Tpost = 4s > τ

• Anchor-Free, Closed-Loop, K = 6, Tpred = Tz* = 4s, Tpost = 0.5s = τ

• Anchor-Based, Open-Loop, K = 2048, Tpred = 4s

• Anchor-Based, Closed-Loop, K = 2048, Tpred = Tz* = Tpost = 4s > τ

• Anchor-Based, Closed-Loop, K = 2048, Tpred = Tz* = 4s, Tpost = 0.5s = τ

• Anchor-Based, Closed-Loop, K = 2048, Tpred = 4s, Tz* = Tpost = 0.5s = τ

• Discrete, Open-Loop, K = 2048, Tpred = 0.5s

• Discrete, Closed-Loop, K = 2048, Tpred = 0.5s

BibTeX


        @misc{lin2025revisitmixturemodelsmultiagent,
          title={Revisit Mixture Models for Multi-Agent Simulation: Experimental Study within a Unified Framework}, 
          author={Longzhong Lin and Xuewu Lin and Kechun Xu and Haojian Lu and Lichao Huang and Rong Xiong and Yue Wang},
          year={2025},
          eprint={2501.17015},
          archivePrefix={arXiv},
          primaryClass={cs.AI},
          url={https://arxiv.org/abs/2501.17015}, 
        }