An integrated safety framework for human-informed learning and injury-aware evaluation of autonomous driving

Autonomous vehicles (AVs) are emerging as intelligent mechanical systems for safer transportation. However, in safety-critical scenarios, existing AVs still rely largely on rule-based, ego-centric strategies, possibly increasing collision and injury risk to other road users. Progress towards safer autonomy is constrained by the rarity of such events and the lack of explicit consideration of post-crash driver injury outcomes. We propose an integrated safety framework for learning and evaluating AV safety performance, enabling AVs to internalize predictable human driving strategies and quantify the extent of injury mitigation in unavoidable collision scenarios. By accelerating the generation of safety-critical behaviors from experienced human drivers and linking pre-crash decisions to in-crash injury through high-biofidelity collision simulation, we evaluate safety performance via human likeness, collision rate, and injury severity. Using 1,988 accelerated driver-in-the-loop collision-avoidance behaviors for model learning and injury-aware evaluation, our human-informed model emulates human-like evasive strategies and, within the experimental paradigm of this study, reduces the combined probability of AIS 3+ injury by 36.6% (95% CI -28.3% to 68.2%) and the median per-crash economic loss by approximately 23% (255,093 USD vs. 196,669 USD) in crashes that still occurred. This work demonstrates that incorporating human evasive experience into autonomous driving learning tends to reduce driver injury severity in collisions under safety-critical scenarios.