Versatile Behavior Diffusion for Generalized Traffic Agent Simulation

Existing traffic simulation models often fail to capture the complexities ofreal-world scenarios, limiting the effective evaluation of autonomous drivingsystems. We introduce Versatile Behavior Diffusion (VBD), a novel trafficscenario generation framework that utilizes diffusion generative models topredict scene-consistent and controllable multi-agent interactions inclosed-loop settings. VBD achieves state-of-the-art performance on the WaymoSim Agents Benchmark and can effectively produce realistic and coherent trafficbehaviors with complex agent interactions under diverse environmentalconditions. Furthermore, VBD offers inference-time scenario editing throughmulti-step refinement guided by behavior priors and model-based optimizationobjectives. This capability allows for controllable multi-agent behaviorgeneration, accommodating a wide range of user requirements across varioustraffic simulation applications. Despite being trained solely on publiclyavailable datasets representing typical traffic conditions, we introduceconflict-prior and game-theoretic guidance approaches that enable the creationof interactive, long-tail safety-critical scenarios, which is essential forcomprehensive testing and validation of autonomous vehicles. Lastly, we providein-depth insights into effective training and inference strategies fordiffusion-based traffic scenario generation models, highlighting best practicesand common pitfalls. Our work significantly advances the ability to simulatecomplex traffic environments, offering a powerful tool for the development andassessment of autonomous driving technologies.

Further reading#

• Access Paper in arXiv.org