溯源 · 规则从何而来
护底不是写出来的。它是挣来的。
安全委员会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? 我们的规则来自 8,219 条关于自动驾驶出错的公开联邦记录:3,336 份 L3–5 系统的 NHTSA 事故报告、4,839 份 L2 报告、21 项自动驾驶软件召回、18 项缺陷调查和 5 份 NTSB 结论。下面每条不变量都指名挣得它的记录。
01 · 记录
五个联邦来源。 下载、校验、一条命令重建。
5,286 条记录至少归入一个失效模式并附带证据。1,830 份 L2 备案完全无法读取:制造商涂黑了叙述并将运动字段留为「未知」。我们如实报告而非隐藏。加州 DMV 自己的报告位于验证码墙后;同一批事故以 State = CA 出现在联邦备案中。
什么算挣得 当至少一条公开记录归入某个提名该不变量的失效模式时,该不变量即为挣得。分类基于规则、可审计:先看结构化事故字段,再看叙述关键词,监管结论显式指定。没有模型阅读记录。
记录中的大多数是什么 L3–5 事故报告中 52% 是自动驾驶车辆停驶时被撞。这不挣得任何规则,并如此计数。护底的规则由系统本身提议了危险的少数事故挣得。
02 · 台账
每条规则由什么挣得。
Records earning each floor invariant, by source Records that earned each invariant, by source one record can earn several invariants · bar length = records · status at right SGO crash reports · ADS SGO crash reports · Level 2 Part 573 recalls ODI investigations NTSB reports 0 250 500 750 1000 1250 ONROAD SGO crash reports · ADS: 167 records → INV_ONROAD SGO crash reports · Level 2: 1,083 records → INV_ONROAD Part 573 recalls: 2 records → INV_ONROAD ODI investigations: 2 records → INV_ONROAD NTSB reports: 1 records → INV_ONROAD 1,255 enforced FOLLOW SGO crash reports · ADS: 538 records → INV_FOLLOW SGO crash reports · Level 2: 646 records → INV_FOLLOW Part 573 recalls: 2 records → INV_FOLLOW ODI investigations: 1 records → INV_FOLLOW 1,187 enforced STATIC SGO crash reports · ADS: 196 records → INV_STATIC SGO crash reports · Level 2: 919 records → INV_STATIC Part 573 recalls: 2 records → INV_STATIC ODI investigations: 2 records → INV_STATIC NTSB reports: 1 records → INV_STATIC 1,120 enforced CROSS_TRAFFIC SGO crash reports · ADS: 546 records → INV_CROSS_TRAFFIC SGO crash reports · Level 2: 495 records → INV_CROSS_TRAFFIC Part 573 recalls: 3 records → INV_CROSS_TRAFFIC ODI investigations: 1 records → INV_CROSS_TRAFFIC NTSB reports: 2 records → INV_CROSS_TRAFFIC 1,047 NOT enforced STOP_FOR_STATIONARY SGO crash reports · ADS: 271 records → INV_STOP_FOR_STATIONARY SGO crash reports · Level 2: 556 records → INV_STOP_FOR_STATIONARY Part 573 recalls: 1 records → INV_STOP_FOR_STATIONARY ODI investigations: 3 records → INV_STOP_FOR_STATIONARY NTSB reports: 3 records → INV_STOP_FOR_STATIONARY 834 enforced VRU SGO crash reports · ADS: 127 records → INV_VRU SGO crash reports · Level 2: 39 records → INV_VRU Part 573 recalls: 1 records → INV_VRU ODI investigations: 1 records → INV_VRU NTSB reports: 1 records → INV_VRU 169 enforced LANE_DISCIPLINE SGO crash reports · ADS: 106 records → INV_LANE_DISCIPLINE SGO crash reports · Level 2: 50 records → INV_LANE_DISCIPLINE Part 573 recalls: 2 records → INV_LANE_DISCIPLINE ODI investigations: 1 records → INV_LANE_DISCIPLINE 159 NOT enforced SPEED_LIMIT SGO crash reports · ADS: 68 records → INV_SPEED_LIMIT Part 573 recalls: 3 records → INV_SPEED_LIMIT ODI investigations: 3 records → INV_SPEED_LIMIT 74 enforced STOP_CONTROL SGO crash reports · ADS: 68 records → INV_STOP_CONTROL Part 573 recalls: 3 records → INV_STOP_CONTROL ODI investigations: 3 records → INV_STOP_CONTROL 74 NOT enforced ODD_ENTRY SGO crash reports · ADS: 44 records → INV_ODD_ENTRY SGO crash reports · Level 2: 25 records → INV_ODD_ENTRY Part 573 recalls: 3 records → INV_ODD_ENTRY ODI investigations: 1 records → INV_ODD_ENTRY 73 NOT enforced WORST_CASE_HEADING SGO crash reports · ADS: 17 records → INV_WORST_CASE_HEADING SGO crash reports · Level 2: 11 records → INV_WORST_CASE_HEADING Part 573 recalls: 2 records → INV_WORST_CASE_HEADING 30 NOT enforced BRAKE_AUTHORITY SGO crash reports · ADS: 1 records → INV_BRAKE_AUTHORITY Part 573 recalls: 2 records → INV_BRAKE_AUTHORITY ODI investigations: 3 records → INV_BRAKE_AUTHORITY NTSB reports: 1 records → INV_BRAKE_AUTHORITY 7 NOT enforced HOLD_AFTER_CONTACT Part 573 recalls: 1 records → INV_HOLD_AFTER_CONTACT ODI investigations: 1 records → INV_HOLD_AFTER_CONTACT 2 NOT enforced MRC_FALLBACK Part 573 recalls: 1 records → INV_MRC_FALLBACK ODI investigations: 1 records → INV_MRC_FALLBACK 2 NOT enforced NO_MOTION_PERSON_ADJACENT Part 573 recalls: 1 records → INV_NO_MOTION_PERSON_ADJACENT 1 NOT enforced
各来源挣得每条不变量的记录数。一条记录可挣得多条不变量。
不变量 规则 记录 重伤 / 致死 执行 ONROAD Stay on the drivable lane/road surface; steer back toward the centre before the edge is reached. 1,255 36 / 6 已执行 STATIC The 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. 1,120 26 / 3 已执行 FOLLOW Keep 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. 1,187 43 / 8 已执行 STOP_FOR_STATIONARY A stationary object or vehicle in the travel lane must be stopped for within the available distance, whatever the lead vehicle just did. 834 26 / 7 已执行 VRU A pedestrian, cyclist, scooter, wheelchair or motorcyclist in or entering the reachable path forces yielding; classification uncertainty never delays braking. 169 31 / 16 已执行 SPEED_LIMIT No throttle above the posted limit. 74 4 / 1 已执行 STOP_CONTROL Complete 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. 74 4 / 1 尚未 CROSS_TRAFFIC Crossing 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. 1,047 31 / 10 尚未 WORST_CASE_HEADING Articulated, towed or ambiguous-orientation objects get a worst-case motion prediction; heading and orientation disagreement widens the reachable set, it never narrows it. 30 2 / 2 尚未 BRAKE_AUTHORITY A 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. 7 0 / 0 尚未 HOLD_AFTER_CONTACT After any collision the vehicle holds still until cleared; no pull-over, no creep. 2 0 / 0 尚未 NO_MOTION_PERSON_ADJACENT No motion from rest while a person is inside the near envelope of the vehicle. 1 0 / 0 尚未 ODD_ENTRY No 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). 73 2 / 1 尚未 LANE_DISCIPLINE Lane changes and merges require a verified gap; never cross the centre line into oncoming traffic. 159 9 / 0 尚未 MRC_FALLBACK On a system fault the vehicle reaches a minimal-risk condition; the floor itself stays live and never hands raw proposals to the actuator. 2 0 / 0 尚未
03 · 欠账
6 条已执行。9 条已挣得但尚未。
下面的网格是本页诚实的核心。绿色单元格给出执行某条规则的护底检查名。红色单元格是记录已经挣得、但两个仿真护底都未执行的规则——仅交叉车流冲突一项就有 1,047 条记录。它们是路线图,以欠账的形式公布,而不是暗示为已覆盖。
Enforcement grid: invariant × floor Which floor enforces which earned invariant ✓ enforced (check name) · ✗ earned by the record, not enforced yet · records = how many earned it records MetaDrive floor CARLA floor ONROAD MetaDrive floor: on_road ✓ on_road CARLA floor: on_road ✓ on_road 1,255 STATIC MetaDrive floor: not enforced ✗ not enforced CARLA floor: on_road(surface guard #0) ✓ on_road + 1,120 FOLLOW MetaDrive floor: following_distance ✓ following_distance CARLA floor: following_distance ✓ following_distance 1,187 STOP_FOR_STATIONARY MetaDrive floor: not enforced ✗ not enforced CARLA floor: emergency_stop ✓ emergency_stop 834 VRU MetaDrive floor: not enforced ✗ not enforced CARLA floor: following_distance/emergency_stop (any actor ahead) ✓ following_distance + 169 SPEED_LIMIT MetaDrive floor: speed_limit ✓ speed_limit CARLA floor: speed_limit ✓ speed_limit 74 STOP_CONTROL MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 74 CROSS_TRAFFIC MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 1,047 WORST_CASE_HEADING MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 30 BRAKE_AUTHORITY MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 7 HOLD_AFTER_CONTACT MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 2 NO_MOTION_PERSON_ADJACENT MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 1 ODD_ENTRY MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 73 LANE_DISCIPLINE MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 159 MRC_FALLBACK MetaDrive floor: not enforced ✗ not enforced CARLA floor: not enforced ✗ not enforced 2
04 · 间距规则里的数字
由实测制动推导, 而非挑选。
记录确定了必须防止什么以及在什么车速下,但没有确定车头时距;物理决定它。停车距离 d = v·tr + v²/(2a)。我们测量了被保护的 MetaDrive 车辆:响应延迟 0.1 s ,持续减速度 11.28 m/s² 。在 55 km/h 场景限速下需要时距 0.78 s ;护底保持 1.6 s。真实汽车制动更弱:若假定 5 m/s²,要求为 1.63 s,1.6 s 便不再足够——这就是 CARLA 护底保持 2.4 s 的原因,也是真实车辆的时距必须由其自身实测制动重新推导的原因。
Stopping distance versus the floor's gap rule Stopping distance the vehicle needs vs the gap the floor holds d_stop(v) = v·t_r + v²/(2a) with measured t_r = 0.1 s; floor gap = max(min_gap, v·headway) 0 m 10 m 20 m 30 m 40 m 50 m 60 m 0 10 20 30 40 50 60 70 80 90 100 110 speed, km/h assumed 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 measured measured 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
05 · 诚实的范围
这证明了什么,没有证明什么。
来源,而非完备。 护底只拒绝它建模的东西。记录和状态中都没有的危险,护底里也没有。
自报的记录。 事故叙述由运营者撰写;计数是下界,不是发生率。
基于规则的分类,抽样审核。 每次匹配附带证据;尚无评审者间一致性研究。
没有危害分析、没有真实感知、没有对抗者,仅限仿真。
来源 sha256 Standing General Order 2021-01, ADS (SAE L3-5) crash reports, Jul 2021 - Jun 15 2025 cb2b38a21e2c… SGO 2021-01 ADS crash reports, Jun 16 2025 onward (third amended order) f856d0b9cedc… SGO 2021-01, Level 2 ADAS crash reports, Jul 2021 - Jun 15 2025 99579d4c9add… SGO 2021-01 Level 2 ADAS crash reports, Jun 16 2025 onward 1b15d76e4b0c… field definitions for the SGO CSVs c92e1bec238e… every Part 573 safety recall since 2010, tab-delimited 306e4fb488c4… every ODI defect investigation (PE/EA/RQ/AQ) 632a7f12837b… Williston FL 2016, Tesla Autopilot vs crossing tractor-trailer (HWY16FH018) 4fca215cec75… Tempe AZ 2018, Uber ATG test vehicle vs pedestrian (HWY18MH010) 1a0bb130efd5… Mountain View CA 2018, Tesla Autopilot into gore attenuator (HWY18FH011) 3784e2ed8219… Delray Beach FL 2019, Tesla Autopilot vs crossing truck (HWY19FH008) 3157beeb98b1… Culver City CA 2018, Tesla Autopilot into stationary fire truck (HWY18FH004) 8a6fc9fd6f2f…
构建时的来源清单(2026-09-27T07:00:26Z)。重建:python -m failure_corpus build。