A Potential Field-Based Approach to Adversarial Inverse Reinforcement Learning for Pedestrian Behavior Modeling

Abstract

End-to-end autonomous driving systems generate control signals directly from sensor inputs, implicitly assuming that the surrounding environment is exogenous. However, at unsignalized intersections, pedestrians actively adjust their actions in response to approaching vehicles, possibly leading to inefficient or unsafe control. To accurately capture interactions across all agents, we propose Mean-Field Adversarial Inverse Reinforcement Learning with Continuous Actions and Kernel Decomposition (MF-AIRL-C+K) to quantitatively evaluate pedestrian reactive behavior in the presence of surrounding pedestrians and vehicles, by extending adversarial inverse reinforcement learning to continuous action spaces and incorporating mean-field-based interaction. Specifically, restricting the utility function to a negative-definite quadratic form guarantees a unique and welldefined Gaussian optimal policy. In addition, the parameterization of the reward function based on kernel decomposition further enforces the condition of potential Mean-Field Games and enables interpretable spatial analysis. Experiments on pedestrian trajectories collected at JR Matsuyama Station show that MF-AIRL-C+K outperforms all baselines on all metrics.