A Multi-Agent Behaviour Generation Algorithm Framework Used for Flexible Historical State Lengths

Simulation is a critical component for assessing autonomous driving technologies. Accurately generating multi-agent behaviours in traffic scenarios is essential for the simulation process. Learning-based approaches have received significant attention in this field. However, these approaches are vulnerable to performance degradation due to distribution shifts when the historical state lengths encountered during testing differ from those used in training-a common occurrence in real-world conditions. Our study introduces a framework for generating multi-agent behaviours in traffic scenarios to address this challenge. This innovative method employs a single model capable of adapting to varying sequence lengths during testing, effectively mitigating the distributional biases stemming from mismatches between test and training historical lengths. The advanced model TrafficFlex, enhanced by our framework, employs knowledge distillation to extract features from past historical states of varying lengths through multiple feature extractors. This approach effectively mitigates issues related to length discrepancies, enabling more accurate behaviour generation across different historical lengths. Through comprehensive testing on the nuScenes dataset, the TrafficFlex has proven its superiority compared to established baselines.