Driver Behavior Modeling with Subjective Risk‐Driven Inverse Reinforcement Learning

This paper proposes a subjective risk‐driven driver behavior modeling approach that incorporates drivers’ risk perception into decision‐making. Inspired by cognitive science, the proposed framework decomposes drivers’ internal evaluation into preference, reward, and subjective risk constraint. The subjective risk is defined as the coupling between drivers’ perceived uncertainty and environmental cost. To obtain a differentiable representation of the subjective risk, we adopt a unified Gaussian potential field formulation that couples the drivers’ cognitive risk fields and environmental cost risk fields through a Gaussian overlap integral. Building upon this, a risk‐threshold Maximum Entropy Inverse Reinforcement Learning paradigm is developed to learn drivers’ internal preference reward and risk perception from demonstrations. Experimental results on large‐scale naturalistic driving data demonstrate that, compared with baseline methods, the proposed approach can stably learn policies that are more consistent with real human driving decision distributions, while maintaining strong performance in high‐risk scenarios. Further human‐in‐the‐loop experiments confirm its effectiveness in personalized modeling, achieving consistency with individual driving behavior across kinematic characteristics, decision‐making patterns, and subjective risk perception. In addition, qualitative analyses show that the learned subjective risk field provides an interpretable representation of drivers’ risk perception, revealing how perceived risk influences driving decision‐making processes.