机器人 / AI 系统 · 溯源研究 · 公开联邦记录

护底从何而来:从公开失效记录中「挣得」运动安全护底的不变量

安全护底的诚实程度,取决于其规则的来源。本文用从 8,219 条公开记录——NHTSA 事故报告、召回与缺陷调查、NTSB 调查结论——中挣得的不变量,取代人工编写的不变量:每条规则引用挣得它的记录,每个间距参数由被保护车辆的实测制动推导。

配套:CARLA 准入研究 → 配套:可准入的运动(理论)→ 溯源研究 · 公开记录 · 可复现构建
每条不变量引用其记录没有模型阅读记录参数来自实测制动
摘要

自动驾驶的运行时保障护底通常由工程师规定、事后论证。我们先回答安全委员会的第一个问题——这些规则从哪里来?——并以机械的方式回答。我们采集了 8,219 条联邦记录:NHTSA《常设通用命令》下 SAE L3–5 系统的 3,336 份事故报告与 L2 系统的 4,839 份(2021 年 7 月至 2026 年 8 月)、21 项自动驾驶软件的 Part 573 召回、18 项缺陷调查办公室(ODI)行动,以及 5 份 NTSB 事故报告。一套确定性、可审计的 23 类失效模式分类法逐条分类:先看结构化字段,再看叙述关键词;没有模型阅读任何记录,每次分类都附带产生它的证据。这些模式共挣得 15 条不变量:6 条已由我们的仿真护底执行;9 条已被记录挣得但尚未执行,其中最大的是交叉车流冲突(1,047 条记录)。L3–5 事故报告中有 52% 描述的是停驶车辆被撞——这不挣得任何规则,并如实报告。间距规则的数值是推导而非选定的:实测响应延迟 0.1 s 与减速度 11.28 m/s² 在 55 km/h 场景限速下要求车头时距 0.78 s,1.6 s 的默认值足以覆盖;若假定真实车辆减速度为 5 m/s²,要求升至 1.63 s,默认值便不再足够。这里的一切都可用一条命令从公开文件重建,并带校验和。

护底不是我们以为危险的东西,而是记录表明已经出过错的东西——写成一条引用其证据的规则。

1 · 安全委员会首先要问的问题

我们此前的论文描述了一个运动运行时保障护底:控制器提议动作,确定性屏蔽层依据一小组不变量验证,只有被准入或被夹紧的动作才到达执行器。那些论文对适用范围是诚实的,但对来源保持沉默。那些不变量——留在路上、遵守限速、保持随速度变化的间距、为前方静止物停车——是工程师写的,其中的数字(1.6 s 车头时距、8 m 最小间距)是构造函数默认值。监管者的第一个问题不是护底在仿真器里是否有效,而是:这些规则从哪里来?你们的失效从哪里采集?

本文以唯一不含糊其辞的方式回答这个问题:从记录出发构建护底的规则集。我们定义:当至少一条公开的自动驾驶失效记录被归入某个提名该不变量的失效模式时,该不变量即为已挣得;并对每条不变量公布:多少条记录挣得它、来自哪些来源、严重程度与车速如何、我们的护底是否已执行。凡规则含有数字之处,该数字都由测量推导,并给出公式。

2 · 记录

五个联邦来源,全部公开,全部以文件形式下载并计算校验和(表 5)。NHTSA《常设通用命令》2021-01 要求每个自动驾驶系统或 L2 驾驶辅助系统的运营者在数日内报告事故;两类系统我们都使用归档文件(2021 年 7 月至 2025 年 6 月 15 日)与当前文件(2025 年 6 月 16 日起),按报告编号保留最高版本。Part 573 召回与 ODI 调查记录来自 NHTSA 的平面文件,按一份显式的自动驾驶活动清单选取;关键词扫描必须精确复现该清单,因此新的备案会导致构建失败而不是被遗漏。NTSB 报告是委员会对自动化状态下事故的五项调查;其「可能原因」段落逐字引用,并在构建时与 PDF 文本核对。

来源记录数
NHTSA SGO 2021-01 事故报告,SAE L3–5(ADS)3,336
NHTSA SGO 2021-01 事故报告,SAE L2(ADAS)4,839
NHTSA 自动驾驶软件 Part 573 召回21
NHTSA ODI 缺陷调查18
NTSB 公路事故报告5
合计8,219

表 1. 记录。5,286 条记录至少归入一个模式;1,830 份 L2 备案因申报方涂黑叙述且运动字段为「未知」而无法分类;1,103 条未匹配任何规则。

有一个缺口是可见而非隐藏的:加州 DMV 的 OL-316 碰撞报告位于验证码墙之后,其网站 API 也不提供。但加州的事件仍然在记录中,因为同一批事故按《常设通用命令》以 State = CA 向联邦申报。

3 · 方法:审阅者可以核查的分类法

失效模式是自动驾驶实际发生过的一种具名失效方式。我们定义了 23 种(图 2)。对事故报告,分类器先读结构化字段——每辆车的碰撞前运动、撞到了什么、哪些面接触、道路标志——再读叙述关键词,且任何违规短语的主语必须是本车(「停在红灯前」是合规而非失效)。召回、调查与 NTSB 报告显式指定模式,因为监管者已经裁定过它们。每次匹配都返回产生它的证据字符串和强度标签:field、text 或 regulator。

定义(已挣得的不变量)。 设 \(R\) 为记录集,\(M\) 为模式集,\(\mathrm{modes}(r)\subseteq M\) 为记录 \(r\) 归入的模式,\(\mathrm{earns}(m)\) 为模式提名的不变量。不变量 \(I\) 为已挣得,当且仅当 \[ E(I) \;=\; \{\, r\in R : \exists\, m\in \mathrm{modes}(r),\; I\in \mathrm{earns}(m) \,\} \;\neq\; \varnothing , \] 且 \(|E(I)|\) 即图 1 报告的计数。一条记录可挣得多条不变量;计数按记录而非按匹配。强证据子集将 \(r\) 限制为字段或监管者证据。

有两种模式按设计不挣得任何规则,但仍然报告。被撞的停驶车辆没有提议任何运动,因此无法由此得出针对本车的规则;未干预的 L2 驾驶员则是关于人类监控的结论,护底是取代它而不是执行它。报告它们使对方责任事故的占比保持可见,而不是让它虚增护底的覆盖面。

Records per failure modeFailure modes found in the record (records per mode)grey = counted but earns no floor rule (stopped vehicle struck; human monitor did not intervene)M01 · Struck while stoppedStruck while stopped: 1,797 records1,797M02 · Contact with static geometryContact with static geometry: 1,078 records1,078M07 · Cross traffic, turns, unprotected leftsCross traffic, turns, unprotected lefts: 772 records772M04 · Front impact while movingFront impact while moving: 726 records726M03 · Lane or road departureLane or road departure: 581 records581M05 · Cut-in absorbed badlyCut-in absorbed badly: 323 records323M22 · Incursion into the ego path by another road userIncursion into the ego path by another road user: 299 records299M23 · Debris or animal in the laneDebris or animal in the lane: 193 records193M08 · Lane change, merge or centre-line violation by theLane change, merge or centre-line violation by the ego: 159 records159M06 · Contact with a vulnerable road userContact with a vulnerable road user: 126 records126M13 · Traffic-control violationTraffic-control violation: 74 records74M14 · Entered a zone the system could not handleEntered a zone the system could not handle: 73 records73M09 · Backing into somethingBacking into something: 47 records47M12 · Mispredicted articulated or towed vehicleMispredicted articulated or towed vehicle: 30 records30M17 · Stationary emergency vehicle in laneStationary emergency vehicle in lane: 15 records15M21 · Human monitor did not interveneHuman monitor did not intervene: 8 records8M10 · Unjustified hard brakingUnjustified hard braking: 6 records6M18 · Crossing vehicle not braked forCrossing vehicle not braked for: 4 records4M11 · Motion after a collisionMotion after a collision: 2 records2M16 · System fault while movingSystem fault while moving: 2 records2M15 · Moved with a person adjacentMoved with a person adjacent: 1 records1M19 · Detected but not acted onDetected but not acted on: 1 records1M20 · Steered into a gore or barrierSteered into a gore or barrier: 1 records1

图 2. 每种失效模式的记录数。灰色模式被计数但不挣得规则。

4 · 结果:记录挣得了什么

Records earning each floor invariant, by sourceRecords that earned each invariant, by sourceone record can earn several invariants · bar length = records · status at rightSGO crash reports · ADSSGO crash reports · Level 2Part 573 recallsODI investigationsNTSB reports025050075010001250ONROADSGO crash reports · ADS: 167 records → INV_ONROADSGO crash reports · Level 2: 1,083 records → INV_ONROADPart 573 recalls: 2 records → INV_ONROADODI investigations: 2 records → INV_ONROADNTSB reports: 1 records → INV_ONROAD1,255enforcedFOLLOWSGO crash reports · ADS: 538 records → INV_FOLLOWSGO crash reports · Level 2: 646 records → INV_FOLLOWPart 573 recalls: 2 records → INV_FOLLOWODI investigations: 1 records → INV_FOLLOW1,187enforcedSTATICSGO crash reports · ADS: 196 records → INV_STATICSGO crash reports · Level 2: 919 records → INV_STATICPart 573 recalls: 2 records → INV_STATICODI investigations: 2 records → INV_STATICNTSB reports: 1 records → INV_STATIC1,120enforcedCROSS_TRAFFICSGO crash reports · ADS: 546 records → INV_CROSS_TRAFFICSGO crash reports · Level 2: 495 records → INV_CROSS_TRAFFICPart 573 recalls: 3 records → INV_CROSS_TRAFFICODI investigations: 1 records → INV_CROSS_TRAFFICNTSB reports: 2 records → INV_CROSS_TRAFFIC1,047NOT enforcedSTOP_FOR_STATIONARYSGO crash reports · ADS: 271 records → INV_STOP_FOR_STATIONARYSGO crash reports · Level 2: 556 records → INV_STOP_FOR_STATIONARYPart 573 recalls: 1 records → INV_STOP_FOR_STATIONARYODI investigations: 3 records → INV_STOP_FOR_STATIONARYNTSB reports: 3 records → INV_STOP_FOR_STATIONARY834enforcedVRUSGO crash reports · ADS: 127 records → INV_VRUSGO crash reports · Level 2: 39 records → INV_VRUPart 573 recalls: 1 records → INV_VRUODI investigations: 1 records → INV_VRUNTSB reports: 1 records → INV_VRU169enforcedLANE_DISCIPLINESGO crash reports · ADS: 106 records → INV_LANE_DISCIPLINESGO crash reports · Level 2: 50 records → INV_LANE_DISCIPLINEPart 573 recalls: 2 records → INV_LANE_DISCIPLINEODI investigations: 1 records → INV_LANE_DISCIPLINE159NOT enforcedSPEED_LIMITSGO crash reports · ADS: 68 records → INV_SPEED_LIMITPart 573 recalls: 3 records → INV_SPEED_LIMITODI investigations: 3 records → INV_SPEED_LIMIT74enforcedSTOP_CONTROLSGO crash reports · ADS: 68 records → INV_STOP_CONTROLPart 573 recalls: 3 records → INV_STOP_CONTROLODI investigations: 3 records → INV_STOP_CONTROL74NOT enforcedODD_ENTRYSGO crash reports · ADS: 44 records → INV_ODD_ENTRYSGO crash reports · Level 2: 25 records → INV_ODD_ENTRYPart 573 recalls: 3 records → INV_ODD_ENTRYODI investigations: 1 records → INV_ODD_ENTRY73NOT enforcedWORST_CASE_HEADINGSGO crash reports · ADS: 17 records → INV_WORST_CASE_HEADINGSGO crash reports · Level 2: 11 records → INV_WORST_CASE_HEADINGPart 573 recalls: 2 records → INV_WORST_CASE_HEADING30NOT enforcedBRAKE_AUTHORITYSGO crash reports · ADS: 1 records → INV_BRAKE_AUTHORITYPart 573 recalls: 2 records → INV_BRAKE_AUTHORITYODI investigations: 3 records → INV_BRAKE_AUTHORITYNTSB reports: 1 records → INV_BRAKE_AUTHORITY7NOT enforcedHOLD_AFTER_CONTACTPart 573 recalls: 1 records → INV_HOLD_AFTER_CONTACTODI investigations: 1 records → INV_HOLD_AFTER_CONTACT2NOT enforcedMRC_FALLBACKPart 573 recalls: 1 records → INV_MRC_FALLBACKODI investigations: 1 records → INV_MRC_FALLBACK2NOT enforcedNO_MOTION_PERSON_ADJACENTPart 573 recalls: 1 records → INV_NO_MOTION_PERSON_ADJACENT1NOT enforced

图 1. 各来源挣得每条不变量的记录数及执行状态。条形为记录计数,挣得两条不变量的记录会出现在两个条形中。

L3–5 事故报告中 52% 是车辆在停驶时被撞。 这一事实重塑了对记录的理解:多数已申报事故并非自动系统提议了危险动作。下面的规则由确实如此的少数事故挣得。

不变量陈述记录数强证据 / 仅叙述重伤 / 致死状态
INV_ONROADStay on the drivable lane/road surface; steer back toward the centre before the edge is reached.1,2551,220 / 3536 / 6已挣得,已执行
INV_STATICThe reachable path must be free of static geometry (pole, gate arm, chain, barrier, curb, wall, attenuator); an unmodelled fixed object in the corridor forces a stop.1,1201,082 / 3826 / 3已挣得,已执行
INV_FOLLOWKeep a speed-dependent gap to whatever is ahead in the lane: gap >= max(min_gap, v * headway), and brake now if the projected gap would fall below min_gap.1,1871,126 / 6143 / 8已挣得,已执行
INV_STOP_FOR_STATIONARYA stationary object or vehicle in the travel lane must be stopped for within the available distance, whatever the lead vehicle just did.834803 / 3126 / 7已挣得,已执行
INV_VRUA pedestrian, cyclist, scooter, wheelchair or motorcyclist in or entering the reachable path forces yielding; classification uncertainty never delays braking.169169 / 031 / 16已挣得,已执行
INV_SPEED_LIMITNo throttle above the posted limit.746 / 684 / 1已挣得,已执行
INV_STOP_CONTROLComplete stop at stop signs and red signals; no entry on a steady yellow without clearance; no passing a stopped school bus with its stop arm out; no through-movement from a turn-only lane.746 / 684 / 1已挣得,未执行
INV_CROSS_TRAFFICCrossing or approaching traffic at intersections and unprotected turns is a threat until its reachable set clears the ego path; a slowly approaching perpendicular vehicle is not assumed to stop.1,0471,047 / 031 / 10已挣得,未执行
INV_WORST_CASE_HEADINGArticulated, towed or ambiguous-orientation objects get a worst-case motion prediction; heading and orientation disagreement widens the reachable set, it never narrows it.302 / 282 / 2已挣得,未执行
INV_BRAKE_AUTHORITYA hard brake must itself be justified by a hazard in the forward reachable set; approach from the rear is not a reason to brake; a justified brake is never suppressed or delayed.76 / 10 / 0已挣得,未执行
INV_HOLD_AFTER_CONTACTAfter any collision the vehicle holds still until cleared; no pull-over, no creep.22 / 00 / 0已挣得,未执行
INV_NO_MOTION_PERSON_ADJACENTNo motion from rest while a person is inside the near envelope of the vehicle.11 / 00 / 0已挣得,未执行
INV_ODD_ENTRYNo entry into a closed zone, standing water, heavy smoke or low visibility; speed is bounded by the distance the sensors can actually see (stop within visible distance).734 / 692 / 1已挣得,未执行
INV_LANE_DISCIPLINELane changes and merges require a verified gap; never cross the centre line into oncoming traffic.159159 / 09 / 0已挣得,未执行
INV_MRC_FALLBACKOn a system fault the vehicle reaches a minimal-risk condition; the floor itself stays live and never hands raw proposals to the actuator.22 / 00 / 0已挣得,未执行

表 2. 不变量台账。「强证据」= 由结构化字段或监管者/NTSB 结论挣得;「仅叙述」= 仅由叙述关键词。严重程度取挣得该项的事故报告中所称的最高伤害。

Enforcement grid: invariant × floorWhich floor enforces which earned invariant✓ enforced (check name) · ✗ earned by the record, not enforced yet · records = how many earned itrecordsMetaDrive floorCARLA floorONROADMetaDrive floor: on_road✓ on_roadCARLA floor: on_road✓ on_road1,255STATICMetaDrive floor: not enforced✗ not enforcedCARLA floor: on_road(surface guard #0)✓ on_road +1,120FOLLOWMetaDrive floor: following_distance✓ following_distanceCARLA floor: following_distance✓ following_distance1,187STOP_FOR_STATIONARYMetaDrive floor: not enforced✗ not enforcedCARLA floor: emergency_stop✓ emergency_stop834VRUMetaDrive floor: not enforced✗ not enforcedCARLA floor: following_distance/emergency_stop (any actor ahead)✓ following_distance +169SPEED_LIMITMetaDrive floor: speed_limit✓ speed_limitCARLA floor: speed_limit✓ speed_limit74STOP_CONTROLMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced74CROSS_TRAFFICMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced1,047WORST_CASE_HEADINGMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced30BRAKE_AUTHORITYMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced7HOLD_AFTER_CONTACTMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced2NO_MOTION_PERSON_ADJACENTMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced1ODD_ENTRYMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced73LANE_DISCIPLINEMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced159MRC_FALLBACKMetaDrive floor: not enforced✗ not enforcedCARLA floor: not enforced✗ not enforced2

图 3. 执行网格。绿色单元格给出执行该不变量的护底检查名;红色单元格是记录已经记下的欠账。

这个网格是本文诚实的核心。6 条不变量已挣得并至少由一个护底执行;9 条已挣得但未执行:STOP_CONTROL、CROSS_TRAFFIC、WORST_CASE_HEADING、BRAKE_AUTHORITY、HOLD_AFTER_CONTACT、NO_MOTION_PERSON_ADJACENT、ODD_ENTRY、LANE_DISCIPLINE、MRC_FALLBACK。仅交叉车流冲突一项就有 1,047 条记录,而两个护底都没有对应检查。碰撞后移动只有两条记录,但其中之一是终结了一家公司无人驾驶运营的 Cruise 召回。

Crash speed envelope per invariantSpeed of the earning crashes, per invariant (mph, moving vehicles only)median ● · 90th percentile ● · maximum ○ — SGO 'SV Precrash Speed'; n = crashes with a recorded speed > 0020406080100120140WORST_CASE_HEADINGmax 82 mphp90 79 mphmedian 38 mphn=24ONROADmax 138 mphp90 75 mphmedian 47 mphn=1101STATICmax 138 mphp90 75 mphmedian 46 mphn=981STOP_FOR_STATIONARYmax 107 mphp90 72 mphmedian 34 mphn=798FOLLOWmax 107 mphp90 69 mphmedian 27 mphn=1134ODD_ENTRYmax 78 mphp90 68 mphmedian 20 mphn=54LANE_DISCIPLINEmax 81 mphp90 67 mphmedian 17 mphn=145CROSS_TRAFFICmax 91 mphp90 62 mphmedian 19 mphn=964SPEED_LIMITmax 62 mphp90 34 mphmedian 12 mphn=66STOP_CONTROLmax 62 mphp90 34 mphmedian 12 mphn=66VRUmax 97 mphp90 21 mphmedian 4 mphn=144

图 4. 每条不变量所对应事故的车速:SGO 碰撞前车速字段的中位数、第 90 百分位与最大值,仅含行驶中的车辆。这是每条规则必须覆盖的包络;它读自记录而非假定。

5 · 间距规则里的数字

记录确定了必须防止什么以及它在什么车速下发生,但没有确定车头时距。这个数字来自被护底保护的车辆的物理特性。以响应延迟 \(t_r\)、持续减速度 \(a\) 从车速 \(v\) 停下所需的距离为

\[ d_{\mathrm{stop}}(v) \;=\; v\,t_r \;+\; \frac{v^{2}}{2a} , \]

而护底的规则对 \(v\) 是线性的:\(\mathrm{gap}(v)=\max(g_{\min},\, v\,h)\)。线性规则恰好覆盖停车距离,当且仅当

\[ h \;\ge\; t_r \;+\; \frac{v}{2a} , \]

因此一个时距在某一车速下恰好精确,在其下方则偏保守。我们在 MetaDrive 车辆上测量了 \(t_r\) 与 \(a\):全油门加速至 30, 45, 55, 70 km/h,然后指令全力制动,每 0.1 s 记录一次车速:延迟 0.1 s(一个控制步长),持续减速度 11.28 m/s²。在 55 km/h 的场景限速下,所需时距为 0.78 s;1.6 s 的默认值有余量地覆盖了它(图 5、表 3)。

Stopping distance versus the floor's gap ruleStopping distance the vehicle needs vs the gap the floor holdsd_stop(v) = v·t_r + v²/(2a) with measured t_r = 0.1 s; floor gap = max(min_gap, v·headway)0 m10 m20 m30 m40 m50 m60 m0102030405060708090100110speed, km/hassumed a = 5 m/s² (not measured)assumed a = 6 m/s² (not measured)assumed a = 8 m/s² (not measured)dashed grey: the same formula at ASSUMED a = 5, 6, 8 m/s² (steepest = 5) — not measuredmeasured a = 11.28 m/s² (MetaDrive)measured a = 11.28 m/s² (MetaDrive)floor gap · 1.6 s / 8 m (MetaDrive default)floor gap · 1.6 s / 8 m (MetaDrive default)floor gap · 2.4 s / 10 m (CARLA)floor gap · 2.4 s / 10 m (CARLA)scenario limit 55 km/h

图 5. 由实测制动(蓝)与假定真实车辆减速度(灰虚线)得到的停车距离,对照两个护底保持的间距。间距线位于停车曲线之上处,规则在该车速下覆盖该车辆。

车速d_stop(实测 a)所需时距护底间距(1.6 s / 8 m)覆盖
20 km/h1.9 m0.35 s8.9 m是
30 km/h3.9 m0.47 s13.3 m是
40 km/h6.6 m0.59 s17.8 m是
50 km/h9.9 m0.72 s22.2 m是
55 km/h11.9 m0.78 s24.4 m是
70 km/h18.7 m0.96 s31.1 m是
90 km/h30.2 m1.21 s40.0 m是
110 km/h44.4 m1.45 s48.9 m是

表 3. 实测 MetaDrive 制动;每一行都是同一公式。

假定 a限速下所需时距限速下 d_stop1.6 s 默认值是否足够
5.0 m/s²1.63 s24.9 m否
6.0 m/s²1.37 s21.0 m是
7.0 m/s²1.19 s18.2 m是
8.0 m/s²1.05 s16.1 m是

表 4. 对假定干路减速度的敏感性——非实测;仿真器制动强于真实车辆。

需要注意的是那些虚线。以 11.28 m/s² 减速的仿真车辆,制动强于干路上的真实汽车。用假定的 6 m/s² 重算同一公式,所需时距为 1.37 s;在 5 m/s² 下为 1.63 s,1.6 s 的默认值便不再足够。CARLA 护底的 2.4 s 覆盖了所有假定行。因此诚实的表述是:MetaDrive 的默认值是推导得出的,对它所保护的车辆足够;对真实车辆,时距必须由该车辆的实测制动重新推导,而 2.4 s 是站得住脚的起点。

6 · 本文确立了什么,没有确立什么

7 · 可复现性

本文的一切,包括图表与正文中的每个数字,都由一次构建从公开文件生成:

python -m failure_corpus build        # fetch → manifest → classify → derive → ledger
python -m failure_corpus.kinematics   # measure MetaDrive braking → kinematics.json
python -m failure_corpus.figures      # the five SVGs
python -m failure_corpus.paper        # this page, EN + 中文
python -m pytest tests/               # per-source, per-mode, per-invariant, per-floor-check pins
来源文件字节sha256
sgo_ads_archiveSGO-2021-01_Incident_Reports_ADS.csv3,943,732cb2b38a21e2ce5c2…
sgo_ads_currentSGO-2021-01_Incident_Reports_ADS.csv2,581,860f856d0b9cedc5f44…
sgo_adas_archiveSGO-2021-01_Incident_Reports_ADAS.csv4,092,53499579d4c9add8f2f…
sgo_adas_currentSGO-2021-01_Incident_Reports_ADAS.csv2,219,1591b15d76e4b0c8bf2…
sgo_dictionarySGO-2021-01_Data_Element_Definitions.pdf529,128c92e1bec238e7578…
recalls_flatFLAT_RCL_POST_2010.zip15,039,755306e4fb488c45e18…
investigations_flatFLAT_INV.zip4,299,580632a7f12837bae96…
ntsb_har1702HAR1702.pdf3,160,7224fca215cec751ee4…
ntsb_har1903HAR1903.pdf2,002,8151a0bb130efd50b99…
ntsb_har2001HAR2001.pdf3,374,2673784e2ed8219b52b…
ntsb_hab2001HAB2001.pdf826,3663157beeb98b18bcd…
ntsb_hab1907HAB1907.pdf811,7888a6fc9fd6f2fa165…

表 5. 构建时的来源清单(2026-09-27T07:00:26Z)。

参考文献

  1. NHTSA. Standing General Order 2021-01 on Crash Reporting, third amended order (2025); incident report files ADS and ADAS, archive 2021–2025 and current. static.nhtsa.gov/odi/ffdd/sgo-2021-01/.
  2. NHTSA Office of Defects Investigation. Recalls flat file FLAT_RCL_POST_2010.zip and investigations flat file FLAT_INV.zip; Part 573 reports 22E-016, 22E-072, 22V-037, 22V-317, 23E-029, 23E-086, 23V-085, 23V-838, 24E-013, 24E-049, 24E-067, 25E-019, 25E-029, 25E-034, 25E-037, 25E-084, 25E-090, 26E-026, 26E-035, 26E-044, 26E-059.
  3. NTSB. HAR-17/02 (Williston, HWY16FH018); HAB-19/07 (Culver City, HWY18FH004); HAR-19/03 (Tempe, HWY18MH010); HAB-20/01 (Delray Beach, HWY19FH008); HAR-20/01 (Mountain View, HWY18FH011).
  4. Sha, L. Using Simplicity to Control Complexity. IEEE Software 18(4), 2001 — the Simplex runtime-assurance architecture.
  5. Shalev-Shwartz, S., Shammah, S., Shashua, A. On a Formal Model of Safe and Scalable Self-driving Cars. arXiv:1708.06374, 2017 — Responsibility-Sensitive Safety; the safe longitudinal distance has the same kinematic form as Section 5.
  6. Alshiekh, M. et al. Safe Reinforcement Learning via Shielding. AAAI 2018.
  7. ASTM F3269-21. Standard Practice for Methods to Safely Bound Behavior of Aircraft Systems Containing Complex Functions Using Run-Time Assurance.
  8. Perslis Research. Runtime Admission Control on a Photoreal Driving Simulator (2026); Admissible Motion (2026). research.perslis.com.