Why some emergency maneuvers succeed: Interpretable behavior graphs from expert drivers to automated control

Emergency handling is among the hardest capabilities for automated driving. Safe evasion demands staged, anticipatory, and tightly coupled longitudinal-lateral actions, yet the expert logic behind such maneuvers is scarce, hard to label by skill, and locked in continuous control signals that obscure why some maneuvers succeed and others fail. This study introduces the Driving Behavior Graph (DBG), an interpretable, probabilistic representation that turns scarce expert demonstrations into reusable knowledge linking human expert decision logic to machine-executable control. Through a two-stage procedure, we screen 25 expert drivers from 99 candidates and collect 400 driver-in-the-loop trials across four emergency scenarios (jaywalking, cut-out, cut-in, merge-in) on a virtual-reality platform. Control signals are segmented with a Hierarchical Dirichlet Process Hidden Semi-Markov Model, mapped to a scenario-specific semantic grid, and assembled into a probabilistic DBG of behavior chains and transitions. The DBG shows that expert emergency behavior is proactive and phased: the first semantic divergence occurred further upstream (semantic bifurcation at longitudinal positions of approximately 55 vs. 115 m in the cut-out scenario), longitudinal and lateral control are coordinated, and the vehicle is deliberately recovered after the conflict; it further localizes the bifurcation points where excellent and poor episodes diverge. Notably, questionnaire-based screening scores were not significantly correlated with objective emergency performance in the present sample, suggesting that screening alone may not fully capture competence in the tested safety-critical scenarios. To close the loop, DBG-derived primitives are synthesized into reference trajectories that guide a model predictive controller, which reproduces expert-like earlier lateral evasion (p < 0.001) and a deliberate safety-over-compliance trade-off, and runs in a vehicle-in-the-loop deployment on a Level-3 automated test vehicle. To support reproducibility, the skill-screened expert emergency-driving dataset is publicly available at https://zenodo.org/records/20710939.