What matters in pre-crash driving behavior modeling? An evaluation of state representations via imitation learning

Investigating pre-crash driving behaviors is vital for uncovering human responses to immediate threats. However, current research mainly relies on naturalistic driving data, capturing only non-critical behavioral patterns. More importantly, the fundamental impact of state representation design on the fidelity of behavioral learning remains largely unexplored. To address these gaps, this study conducts systematic state representation engineering for pre-crash scenarios. We formulate four progressive state representations, evolving from basic geometric coordinates to a comprehensive risk-aware formulation. To evaluate their impact, these representations are integrated into three distinct imitation learning (IL) algorithms to model behavioral patterns from real-world pre-crash trajectory data. The trained policies are subsequently evaluated in a simulated interactive environment: the agent controls the ego vehicle to interact with an object vehicle that replays testing-set trajectories, allowing for a direct comparison of maneuvers against authentic human responses. Experimental results demonstrate that while algorithmic selection influences performance, the state representation plays also a significantly decisive role. Across all three IL algorithms, agents utilizing the risk-aware representation consistently generate maneuvers most similar to those of the real-world ego vehicle. Furthermore, an analysis of the extracted behavioral rewards indicates that the risk-aware representation enables the agent to internalize human driving logic, moving beyond the mere mechanical replication of observed trajectories.