AR2G-AIRL: A Context-Adaptive Game-Inverse Reinforcement Learning Algorithm for Modeling Pedestrian--Vehicle Interaction Considering Group Behaviors

Pedestrian-vehicle interactions are critical problems in the development of autonomous driving technologies. However, there is still a lack of comprehensive understanding of pedestrian behaviors during pedestrian-vehicle interactions, especially considering the impact of group behaviors. This study proposes an Actor-Relation-Graph-based Game-Adversarial Inverse Reinforcement Learning (AR2G-AIRL) algorithm to understand pedestrian behaviors during pedestrian-vehicle interactions, considering the influence of surrounding pedestrians and vehicles. Specifically, we first identify pedestrian-vehicle interactions from a naturalistic driving dataset as the data basis. Subsequently, the AR2G-AIRL algorithm is constructed with three components: Actor-Relation Graphs (ARGs) to capture the features within groups, evolutionary game theory to extract the features between groups, and the Adversarial Inverse Reinforcement Learning (AIRL) algorithm to recover the reward function. The experiment results illustrate that the AR2G-AIRL algorithm can accurately simulate pedestrian behaviors during pedestrian-vehicle interactions and outperform all the baseline methods and ablation studies. Meanwhile, the trajectory and ARG weight visualization results demonstrate the importance of surrounding pedestrians under diverse interactions. Moreover, the analysis of the recovered reward function validates the “safety in numbers” effect and reveals the basic patterns of pedestrians interacting with vehicles without surrounding pedestrians and the heterogeneous impact of the size and morphologies of pedestrian groups on pedestrian behaviors. These results highlight the importance of incorporating group behaviors, offering insights into enhancing road safety and the efficiency of autonomous systems.