一句话：提出用离线峰值转向偏差排序相机扰动族，从而在有限车载测试预算内更早发现 ADS 失败。
论文关注 ADS 车载测试排序：扰动录制相机图像可复现，但缺少转向、车辆运动与后续图像间的闭环反馈；车载测试能捕获闭环，却依赖车辆、路线与安全监督。因测试活动可能提前停止，固定运行预算下顺序决定暴露哪些失败。作者回顾性分析 40 种相机扰动、每种 5 个强度，针对 DAVE-2 风格 ADS，并复用协议：先测最大强度，失败后逐级下降至首次成功。按离线峰值转向偏差排序扰动族后，前 10 次最大强度运行均发现失败（随机期望 4.0），失败发现 AUC 从 0.513 升至 0.756；全部 77 次运行后优势仍存（0.618 对 0.522）。离线信号不完美，不能精确预测失败或跳过强度，但与更早发现失败相关。
Perturbing recorded camera images is a reproducible way to test an autonomous driving system (ADS), but it omits the feedback between steering, vehicle motion, and subsequent images. On-vehicle tests capture this closed loop but require a vehicle, route, and safety supervision. Because a campaign may stop before every candidate is executed, test order determines which failures are exposed within a fixed run-count budget. We therefore ask whether inexpensive offline measurements can prioritize which perturbation family to test next on a vehicle. We retrospectively analyze 40 camera perturbations, each configured at five intensities, for a DAVE-2-style ADS. The reused protocol tests every family at maximum intensity and, after failure, descends until the first observed success. Ranking families by peak offline steering deviation finds failures in all first 10 maximum-intensity runs, compared with 4.0 expected under random ordering, and raises failure-discovery AUC from 0.513 to 0.756. Counting all 77 protocol runs reduces but preserves the advantage (0.618 versus 0.522). Thus, in this campaign, an imperfect offline signal was associated with earlier failure discovery without serving as an exact failure predictor or justifying skipped intensities.