AutoZYX
Remote Operations & Anomaly Management
L4 自动驾驶远程运营与异常管理开源框架事件数据库 · 场景分类 · 10 种运营模式 · 参考架构 · 评价基准 · 技术支柱
Open-source incident database, scenario taxonomy,
reference architecture & evaluation benchmarks for robotaxi remote operations
v1.2 · 最后更新：2026-05-06Updated: 2026-05-06 · 已对齐加州 DMV Article 3.7（Express Terms 2026）Aligned with California DMV Article 3.7 (Express Terms 2026)
为什么需要这个项目？
• 萝卜快跑：2026年3月31日，武汉近百辆萝卜快跑在高架路段同时熄火，乘客被困2小时，SOS无人应答，最终依靠交警逐车施救。事件后14天内，国家三部门召开全国安全监管会议，超10个试点城市同步收紧准入，行业监管逻辑从"包容审慎"转向"安全底线刚性化"。
• Waymo：三个月前，旧金山停电导致Waymo大面积瘫痪，应急热线等待2.5小时。
• 美国国会调查：2026年3月，参议员Markey发布《Remote Back Seat Operators》报告，七家头部自动驾驶公司（Aurora、May Mobility、Motional、Nuro、Tesla、Waymo、Zoox）均拒绝披露其无人车需要远程人工干预的频率。
• 加州 DMV Article 3.7：2026 年 Express Terms 修订版正式确立 "Remote Operations Support" 为法定术语，区分 Remote Driver 与 Remote Assistant 两类岗位，规定 14 项强制安全档案要素、30 秒紧急响应、§ 227.46 分级运营限制等可量化阈值——全球首部远程运营专门法规。
• 当无人驾驶出租车规模化运营时，这种靠人力逐车施救的模式不可持续，而行业连最基本的透明度都无法保证。Why this project?
• Baidu Apollo: On March 31, 2026, ~100 robotaxis shut down simultaneously on Wuhan's elevated highways. Passengers trapped for 2 hours. SOS unanswered. Police rescued vehicles one by one.
• Waymo: Three months earlier, a San Francisco blackout stranded Waymo's fleet — emergency services waited 2.5 hours on hold.
• US Congress: In March 2026, Senator Markey's investigation revealed that all seven major AV companies (Aurora, May Mobility, Motional, Nuro, Tesla, Waymo, Zoox) refused to disclose how frequently their vehicles require remote human intervention.
• California DMV Article 3.7: The 2026 Express Terms amendment officially codified "Remote Operations Support" as a legal term, distinguished Remote Driver vs. Remote Assistant roles, and prescribed 14 mandatory safety case elements, 30-second emergency response, § 227.46 graduated restrictions, and other quantifiable thresholds — the first dedicated regulation worldwide for AV remote operations.
• At scale, manual vehicle-by-vehicle rescue is unsustainable — and the industry cannot even guarantee basic transparency.
ROAM 覆盖 10 种 L4 自动驾驶运营模式，涵盖商业运营、个人所有、公共交通、专用环境四大类ROAM covers 10 L4 operating modes across commercial, personal ownership, public transit, and specialized environments
关键洞察：L4 的未来不是单一的 Robotaxi。随着技术成熟，OEM 会将 L4 能力销售给消费者（类 Tesla 模式）；车主可能通过订阅方式使用 L4 能力；车辆闲时可能自主去做出租运营。此时 OEM 从制造商升级为服务提供者，必须为每一辆售出的 L4 车辆提供 7×24 远程运营支持。ROAM 正是为这种多模式未来设计的开源治理框架。Key insight: The future of L4 is not only Robotaxi. As technology matures, OEMs will sell L4 capability to consumers (Tesla-like model); owners may subscribe to L4 services; idle vehicles may autonomously operate as ride-hail. At that point OEMs transition from manufacturers to service providers, obliged to provide 24/7 remote operations support for every L4 vehicle sold. ROAM is designed as the open governance framework for this multi-mode future.
| 类别Category | 模式Mode | 车辆所有Ownership | 远程运营责任Remote Ops Responsibility | 代表Examples |
|---|---|---|---|---|
| 商业运营Commercial | M1 Robotaxi 运营商车队Robotaxi Operator Fleet | 运营商Operator | 运营商 ROCOperator ROC | Waymo · 滴滴 · Apollo Go · Pony.ai |
| M2 OEM 直营车队OEM Direct Fleet | OEM | OEM ROC | Tesla Robotaxi · Zeekr Mobility | |
| M3 车队即服务Fleet-as-a-Service | 租赁商Leasing Co. | 平台+车队主Platform + Fleet owner | Hertz × AV | |
| 个人所有Personal | M4 个人所有 + OEM 订阅Personal + OEM Subscription | 消费者Consumer | OEM（服务化）OEM (service model) | Tesla FSD 订阅 · 蔚来 |
| M5 个人所有 + 闲时共享Personal + Shared | 消费者Consumer | OEM/平台动态切换OEM/Platform dynamic | Tesla Robotaxi Network | |
| M6 纯私人 L4Pure Private L4 | 消费者Consumer | OEM+车主兜底OEM + owner fallback | Mercedes DRIVE PILOT | |
| 公共交通Public Transit | M7 L4 公交巴士L4 Public Bus | 公交/政府Transit/Gov | 双责任Dual responsibility | 轻舟 Robobus · 宇通 |
| M8 L4 微循环接驳Micro-Transit | 运营方Operator | 运营方Operator | 园区/景区接驳Campus/park shuttle | |
| 专用环境Specialized | M9 无人配送Last-Mile Delivery | 物流企业Logistics | 物流企业Logistics ROC | 美团 · 京东 · Nuro |
| M10 特定环境专用Restricted Environment | 客户企业Client | 集成商+客户Integrator+client | 矿区/港口/机场Mine/port/airport |
新增异常类别：在 M5 模式（个人所有 + 闲时共享）下，存在全新的"所有权/使用权切换冲突"异常——例如车主紧急召回正在载客的车辆、共享期间的保险责任边界等。ROAM 场景分类 v1.1 将新增 Category G 专门覆盖这类异常。Novel anomaly class: M5 (Personal + Shared) introduces a new "ownership/usage transition conflict" anomaly type — e.g., owner urgently recalling a vehicle in service, insurance boundary during shared sessions. Scenario Taxonomy v1.1 will add Category G to cover these.
完整定义：operating-models-v1.0.md · 责任矩阵：responsibility-matrix.md Full definitions: operating-models-v1.0.md · Responsibility matrix: responsibility-matrix.md
构成 L4 自动驾驶远程运营的完整开源基线A complete open-source baseline for L4 autonomous mobility remote operations
580+ 结构化事件记录，覆盖 Waymo、Cruise、百度 Apollo、Pony.ai 等 14+ 家运营商（2019–2026），YAML 标准格式，社区持续更新580+ structured records across 14+ operators (Waymo, Cruise, Baidu Apollo, Pony.ai, etc., 2019–2026), standardized YAML format, community-maintained
6大类、29子场景（含 v1.1 新增 E5 紧急地理围栏 / E6 第一响应人 Override 系统），含严重度 S0–S4 和紧急度 U0–U3 分级，对齐 ISO 34502 场景分层方法6 categories, 29 sub-scenarios (incl. v1.1-new E5 emergency geofencing / E6 first-responder override), with severity S0–S4 and urgency U0–U3 scales, aligned with ISO 34502
三层决策模型 + 责任矩阵 v1.2（10 模式 × 3 层 × 7 维度，含 v1.2 新增"监管反馈维度"）+ FRIP 模板（紧急响应人员交互计划）Three-layer decision model + Responsibility matrix v1.2 (10 modes × 3 layers × 7 dimensions, v1.2 adds "regulatory feedback") + FRIP template (First Responder Interaction Plan)
16 项 KPI：8 项基线（MTTR、AI 处置率等）+ v1.1 新增 5 项法规对齐（KPI 9-13，含 30s 紧急响应阈值）+ v1.2 新增 3 项监管反馈（KPI 14-16，含限制合规时限、解除周期、复发率）16 KPIs: 8 baseline (MTTR, AI resolution rate, etc.) + v1.1 5 regulator-aligned (KPIs 9-13, incl. 30s emergency response threshold) + v1.2 3 regulatory feedback (KPIs 14-16: restriction compliance time, lift cycle, recurrence rate)
AI优先处置，人工逐级兜底AI-first response, human fallback by escalation
场景识别 → 风险评估 → 决策生成 → 自动执行，无需人工介入Scenario recognition → risk assessment → decision generation → autonomous execution
靠边停车 · 驶入服务区 · 低速绕障 · 感知模块重启 · 乘客语音安抚 · 调度备用车辆Pull over · enter service area · low-speed reroute · perception restart · passenger voice guidance · dispatch backup
↳ GB标准C.2.1.2：ADS应不依赖远程协助执行DDT↳ GB C.2.1.2: ADS shall not depend on remote assistance for DDT
AI生成处置方案（含置信度）→ 远程安全员审核确认或修改 → 车辆执行AI generates action plan with confidence → remote operator reviews/confirms → vehicle executes
复杂路口脱困 · 交警手势解读 · 多车冲突裁决 · 报警/叫救护车决策 · 乘客异常行为Complex intersection escape · police gesture interpretation · multi-vehicle conflict · emergency call decision
↳ GB标准C.2.3：ADS应能够及时接收远程协助指令↳ GB C.2.3: ADS shall receive remote assistance commands in a timely manner
远程人员直接接管车辆控制（≤10km/h）或派遣线下救援团队到现场Remote operator takes direct vehicle control (≤10km/h) or dispatches on-site rescue team
车辆硬件故障 · 批量系统瘫痪 · 事故现场处理 · 车辆起火 · 乘客物理救援Hardware failure · mass system shutdown · accident scene · vehicle fire · passenger physical rescue
↳ 德国StVFernLV远程驾驶框架 / Tesla Teleops模式↳ German StVFernLV / Tesla Teleops model
基于Koopman六角色框架，明确远程运营中每类人员的职责边界Based on Koopman's six-role framework, clarifying responsibility boundaries of each personnel type
超越"Assistant vs Driver"的二分法。 当前行业对远程协助人员的定位存在根本分歧：他们是"辅助者"还是"驾驶员"？Phil Koopman在2026年4月提出六角色框架，将人类在自动驾驶系统中的参与细分为6种职责×2种位置（现场/远程）=12种角色组合，彻底重构了这一讨论。Beyond the "assistant vs. driver" dichotomy. The industry fundamentally disagrees on how to classify remote operators: assistants or drivers? In April 2026, Phil Koopman proposed a six-role framework that breaks human involvement into 6 responsibilities × 2 locations (on-site/remote) = 12 role combinations, restructuring this debate.
| 角色Role | 职责Responsibility | 对应ROAM层级ROAM Layer | 安全责任Safety Responsibility |
|---|---|---|---|
| 操作员 OperatorOperator | 执行持续控制回路（方向、制动）Performs continuous control loop (steering, braking) | Layer 3 | 完整驾驶员责任Full driver liability |
| 监督员 SupervisorSupervisor | 主动监控自动化，必要时介入Actively monitors automation, intervenes proactively | 跨层级 | 介入响应责任Intervention responsibility |
| 决策者 DeciderDecider | 响应AI请求，批准或修改方案Responds to AI queries, approves/rejects plans | Layer 2 | 决策范围责任Decision-scope liability |
| 响应者 ResponderResponder | 处理异常事件（脱困、事故、乘客问题）Handles incidents (extrication, accidents, passenger issues) | Layer 3 | 情景判断责任Situational judgment |
| 待命者 StandbyStandby | 按需切换至任何其他角色On-call to switch into any role | 调度层Orchestration | 就绪与警觉Availability and alertness |
| 维护者 MaintainerMaintainer | 诊断、维修、运营就绪Diagnostics, repair, operational readiness | 正交维度Orthogonal | 维护质量Maintenance quality |
关键洞察：Layer 2的远程决策者（Remote Decider）并非单纯的"辅助者"——他们对交通信号灯识别、道路物体分类、路径批准的决策直接影响ADS行为，这些本就是DDT的组成部分。Koopman明确反对"责任洗白"（accountability laundering）——将安全责任转移给AI计算机是回避人类问责的不当做法。Key insight: Layer 2 Remote Deciders are NOT merely "assistants" — their decisions on traffic signal interpretation, object classification, and path approval directly affect ADS behavior and are legitimate DDT components. Koopman explicitly rejects "accountability laundering" — shifting safety responsibility to AI computers is an improper evasion of human accountability.
参考：Koopman, P., "Human Roles in AVs and Embodied AI Systems," 2026-04-04. 原文链接 Reference: Koopman, P., "Human Roles in AVs and Embodied AI Systems," April 4, 2026. Original article
借鉴美军UAV与DJI远程无人机遥控产品化沉淀，避免重走弯路Military UAV and DJI remote drone operations productization, avoiding repeated mistakes
自动驾驶行业正在重复无人机领域30年前犯过的错误。Cummings（2026）指出美军UAV远程运营的五大教训可直接迁移至Robotaxi；而DJI已将这些教训工程化为成熟产品。卓驭科技（承袭DJI技术栈）在低延迟图传、操作员工作站、飞手认证等维度具备原生优势。The self-driving industry is repeating mistakes drone operators made 30 years ago. Cummings (2026) identified five military UAV lessons directly applicable to Robotaxi; DJI has already productized these lessons. Zhuoyu Technology (inheriting DJI's stack) holds native advantages in low-latency transmission, operator workstations, and pilot certification.
| 教训Lesson | 军方经验（Cummings）Military Experience | DJI工程解答DJI Engineering Answer |
|---|---|---|
| 延迟约束Latency | 人类神经肌肉延迟200-500ms；2秒延迟导致起降事故率极高200-500ms human neuromuscular lag; 2s latency caused high UAV crash rate | OcuSync/O4物理层直连，端到端延迟低至15ms，15km范围自适应跳频OcuSync/O4 physical-layer link, 15ms latency, 15km range with adaptive hopping |
| 工作站设计Workstation | Shadow UAV 80%事故因界面缺陷；Global Hawk仅损失1架（界面优良）Shadow UAV: 80% of crashes from poor UI; Global Hawk lost only 1 | FlightHub 2 Virtual Cockpit三段式界面+力反馈控制器FlightHub 2 Virtual Cockpit: three-panel UI + force-feedback controller |
| 操作员负荷Operator Load | Waymo 1:29人车比已达天花板；2024-2025响应时间无改善Waymo 1:29 ratio at capability ceiling; no improvement 2024-2025 | Dock 2本地自主+云端协同，操作员仅处理异常事件Dock 2 local autonomy + cloud coordination, operators handle exceptions only |
| 训练认证Training | 自动驾驶行业无KSA标准、无认证checkrideAV industry has no KSA standards, no certification checkride | DJI UTC全球200+培训中心，40,000+认证飞手DJI UTC: 200+ global training centers, 40,000+ certified pilots |
| 应急预案Contingency | Waymo断电时无法召回车辆；常态配置在应急时崩溃Waymo couldn't recall vehicles during blackout; nominal staffing fails in emergencies | Dock 2：IP55+双RTK+电池备份+通信失联自动返航Dock 2: IP55 + dual RTK + battery backup + auto return-to-dock |
ROAM项目的独特贡献：将DJI/卓驭的无人机远程运营技术栈系统化提炼为Robotaxi远程运营参考标准，包括专用低延迟传输链路、操作员工作站规范、飞手级认证体系。这是其他纯自动驾驶背景公司难以复制的结构性优势。ROAM's unique contribution: systematically distilling DJI/Zhuoyu's drone remote operations stack into Robotaxi reference standards — dedicated low-latency transmission, operator workstation specifications, pilot-grade certification. This is a structural advantage other pure-AV companies cannot easily replicate.
详细分析：uav-lessons-learned.md · 参考：Cummings (2026), DJI Enterprise Full analysis: uav-lessons-learned.md · References: Cummings (2026), DJI Enterprise
2026年3月，参议员Markey发布首份针对自动驾驶远程协助操作的国会调查报告March 2026: Senator Markey published the first Congressional investigation into AV remote assistance operations
七家头部自动驾驶公司（Aurora、May Mobility、Motional、Nuro、Tesla、Waymo、Zoox）均拒绝披露其无人车需要远程人工干预的频率——这正是ROAM开放事件数据库旨在填补的透明度空白。Seven major AV companies (Aurora, May Mobility, Motional, Nuro, Tesla, Waymo, Zoox) all refused to disclose how frequently their AVs require remote human intervention — precisely the transparency gap ROAM's open incident database aims to fill.
| 公司Company | 平均延迟Avg Latency | 最差Worst |
|---|---|---|
| May Mobility | 100–140 ms | 500 ms |
| Waymo （本土）(domestic) | 150 ms | — |
| Nuro | 200 ms | — |
| Waymo （海外）(overseas) | 250 ms | — |
| Motional | 80–300 ms | — |
| Zoox | — | 333 ms |
| Aurora / Tesla | 拒绝披露Refused to disclose | |
对ROAM的意义：Markey报告从立法层面印证了ROAM项目的核心假设——行业在远程运营透明度方面存在系统性缺失，需要独立的开放数据基础设施来填补。报告中首次披露的跨公司延迟数据也为ROAM的KPI基准体系提供了实证参考。Implications for ROAM: The Markey report legislatively validates ROAM's core thesis — the industry suffers from systemic opacity in remote operations, requiring independent open data infrastructure. The first-ever cross-company latency data also provides empirical benchmarks for ROAM's KPI framework.
报告全文：Markey RAOS Report (PDF) · 中文解读：智车科技分析 Full report: Markey RAOS Report (PDF) · Chinese analysis: Smart Car Technology article
2026 年 Express Terms 修订版与 ROAM 框架在术语、岗位分工、量化阈值、紧急响应等四个维度全方位对齐2026 Express Terms aligned with ROAM v1.2 across terminology, role separation, quantitative thresholds, and emergency response
Markey 报告揭示了行业的透明度缺失，加州 DMV 紧随其后给出了监管答案。Article 3.7（Express Terms 2026）是全球首部以"远程运营"为专门对象的法规，正式确立 "Remote Operations Support" 为法定术语，明确区分 Remote Driver（实时控制 / 9 项功能要求）与 Remote Assistant（提供信息建议 / 7 项功能要求），并规定 14 项强制安全档案要素与多项可量化阈值。ROAM v1.2 已系统对齐这一法规体系。Where Markey exposed the industry's transparency gap, California DMV provided the regulatory answer. Article 3.7 (Express Terms 2026) is the first regulation worldwide dedicated to AV remote operations: it codifies "Remote Operations Support" as a legal term, distinguishes Remote Driver (real-time control / 9 functional requirements) from Remote Assistant (information & advisory / 7 functional requirements), and prescribes 14 mandatory safety case elements plus quantitative thresholds. ROAM v1.2 has been systematically aligned with this regulatory framework.
DMV § 227.04 要求每家 driverless 测试 / 部署许可申请方提交涵盖 ODD、运营设计、网络安全、远程运营、事故应对等 14 项强制要素的安全档案。ROAM 四个模块（事件库 / 场景分类 / 参考架构 / 评价基准）逐项映射，可作为企业起草安全档案的开源工作底稿。DMV § 227.04 requires every driverless testing/deployment applicant to submit a safety case covering 14 mandatory elements (ODD, operational design, cybersecurity, remote operations, incident response, etc.). ROAM's four modules map item-by-item, serving as an open-source working baseline for safety case drafting.
DMV § 227.38 定义 Remote Driver（实时控制车辆，承担完整驾驶员责任，对应 ROAM Layer 3）；§ 227.40 定义 Remote Assistant（仅提供信息和建议，对应 ROAM Layer 2 AI 辅助 + 人工确认）。ROAM 三层决策模型在法规层面得到验证。DMV § 227.38 defines Remote Driver (real-time control with full driver liability, mapped to ROAM Layer 3); § 227.40 defines Remote Assistant (information and advisory only, mapped to ROAM Layer 2 AI-assisted + human confirm). ROAM's three-layer model is validated at the regulatory level.
DMV § 227.46 在"无处罚"和"完全吊销"之间引入 5 档处置（整改通知 → 日常运营限制 → 强制人员介入 → 立即限制 → 完全吊销）。ROAM 责任矩阵 v1.2 新增"监管反馈维度"作为第七维度，对应 KPI 14（合规时限 ≤24h / 紧急 ≤1h）、KPI 15（解除申请周期）、KPI 16（复发率 <10%）。DMV § 227.46 introduces 5 graduated tiers between "no action" and "full revocation" (rectification notice → daily restriction → mandatory personnel → imminent hazard → revocation). ROAM Matrix v1.2 adds "regulatory feedback dimension" as the 7th outer-layer dimension, mapped to KPI 14 (compliance time ≤24h / imminent ≤1h), KPI 15 (lift application cycle), KPI 16 (recurrence rate <10%).
DMV § 227.42(i) 强制要求 driverless 申请方提交涵盖 15 项法定要素的 First Responder Interaction Plan（含 ODD、紧急联系电话、安全接近 / Override 程序、车辆移动协调、灭火、拖车等）。ROAM v1.2 新增完整模板，含 15 节主体 + 重型车扩展节，每节含法规依据、内容模板和 ROAM 框架对应关系。DMV § 227.42(i) requires driverless applicants to submit a First Responder Interaction Plan covering 15 mandatory elements (ODD, emergency phone, safe approach/override, vehicle movement coordination, fire, towing, etc.). ROAM v1.2 introduces a complete template with 15 main sections plus a heavy-duty extension, each section providing legal citation, content template, and ROAM framework cross-reference.
对中国标准化工作的启示：加州 DMV 的术语体系与量化阈值如果率先成为事实标准，中国标准化工作可能面临"被加州术语裹挟"的风险。当前更现实的路径是先通过 CSAE / CAICV 团标预研形成中国本土共识：在术语层面参照但不照搬 DMV 二分法，结合《自动驾驶汽车运输安全服务指南》和 YD/T 4778-4783 形成本地化定义；在量化阈值上对齐 30s/2min 等基线，但根据中国实际场景（如 1:3 安全员比例）适度调整。Implications for Chinese standardization: If California DMV's terminology and quantitative thresholds become the de facto global standard, Chinese standards may risk being "captured by California terminology". The more practical near-term path is to build Chinese-context consensus through the CSAE / CAICV group-standard pre-research: reference but do not copy DMV's binary classification, combine the Chinese MoT "AV Transport Safety Service Guidelines" and YD/T 4778-4783 to form localized definitions, and align on 30s/2min baselines while adjusting for China-specific contexts such as the 1:3 safety-operator ratio.
完整映射文档：california-article-3-7-mapping.md（H 章节体系映射 + 14 项安全档案对接 + 9 大关键术语） · 法规原文：DMV Article 3.7 Modified Text PDF Full mapping doc: mapping document (8-chapter mapping + 14 safety case items + 9 key terms) · Original text: DMV Article 3.7 Modified Text PDF
L4 远程运营需要多个技术领域深度融合，单一技术无法独立支撑L4 remote operations require deep integration of multiple technology domains — no single technology can stand alone
异常自识别、不确定性量化、ODD 监控、SOTIF 触发条件。决定"异常能否被识别"。Anomaly self-detection, uncertainty quantification, ODD monitoring, SOTIF triggers. Determines "can anomalies be detected".
直接/间接控制、waypoint 规划、操作员情境感知、认知负荷管理。玻璃到玻璃延迟 < 200 ms。Direct/indirect control, waypoint planning, operator situation awareness, cognitive load management. Glass-to-glass latency < 200 ms.
5G URLLC、网络切片、C-V2X、5G-Advanced、LEO 卫星备份。99.999% 可靠性，抖动 < 5 ms。5G URLLC, network slicing, C-V2X, 5G-Advanced, LEO satellite backup. 99.999% reliability, jitter < 5 ms.
决策置信度、自动升级机制、LLM 增强、可信 AI 治理。信任校准误差 < 15%。Decision confidence, auto-escalation, LLM augmentation, trustworthy AI governance. Trust calibration error < 15%.
整车安全、OTA 安全、数据主权、隐私保护、远程通道安全。CVE 补丁 < 30 天。Vehicle security, OTA security, data sovereignty, privacy, remote channel security. CVE patch < 30 days.
操作员资质、训练认证、工作负荷、跨时区协同、应急组织。班次 ≤ 2 h，初训 ≥ 120 h。Operator qualification, training, workload, cross-timezone coordination, emergency organization. Shift ≤ 2 h, initial training ≥ 120 h.
中国标准与国际标准的双语桥接：六大支柱涉及的中国标准与国际标准大多存在对应关系：ISO 21448↔GB/T 43267-2023（SOTIF）、ISO/SAE 21434↔GB 44495（网络安全）、UN R156↔GB 44496（OTA）、SAE J3016↔GB/T 40429（自动驾驶分级）、3GPP TS 22.186 Remote Driving↔YD/T 4778-4783（5G 远程遥控驾驶）、ISO 34502↔GB/T 41798（场景测试）。这为国内研究者提供了对标国际前沿的明确路径。China-International Standards Bridge: Most technology pillar standards have Chinese counterparts: ISO 21448↔GB/T 43267-2023 (SOTIF), ISO/SAE 21434↔GB 44495 (cybersecurity), UN R156↔GB 44496 (OTA), SAE J3016↔GB/T 40429 (automation levels), 3GPP TS 22.186 Remote Driving↔YD/T 4778-4783 (5G tele-driving series), ISO 34502↔GB/T 41798 (scenarios). This provides Chinese researchers with clear alignment to international frontier.
完整参考文献与行业最佳实践：technology-pillars.md · 50+ 学术文献、20+ 标准、15+ 行业案例 Full references and best practices: technology-pillars.md · 50+ academic papers, 20+ standards, 15+ industry cases
全球尚无专门的 L4 远程运营与异常管理参考架构标准 —— ROAM 旨在填补这一空白No international standard specifically addresses L4 remote operations & anomaly management — ROAM aims to fill this gap
经系统扫描（2026 年 4 月），目前国际上没有任何一部 ISO/SAE/IEC 标准以"L4 远程运营与异常管理"为专门对象。所有现有相关标准都是"要求型"（requirements-oriented）文档，无一是"参考架构型"（reference architecture）。SAE J3016 明确拒绝定义 teleoperation；德国 StVFernLV 和英国 AV Act 2024 是法规而非技术标准。这是 ROAM 的独占性战略高地。After systematic scanning (April 2026), no ISO/SAE/IEC standard specifically addresses "L4 remote operations and anomaly management". All existing standards are requirements-oriented, none provide a reference architecture. SAE J3016 explicitly refuses to define teleoperation; Germany StVFernLV and UK AV Act 2024 are regulations, not technical standards. This is ROAM's exclusive strategic opportunity.
战略节奏：Strategic Pace: ROAM 继续作为开源研究框架沉淀公开事件、场景分类、责任矩阵和 KPI；CSAE 团标作为可标准化子集推进中国本土共识。当前标准题名为《智能网联汽车 高度自动驾驶车辆运营管理异常事件处置规范》，仍处于预研和起草组征集阶段，尚非正式发布标准。ROAM continues as the open-source research framework for public incidents, scenario taxonomy, responsibility matrices, and KPIs; the CSAE group standard is the standardizable Chinese-context subset. The current proposed title is "Specification for Operations Anomaly Event Management of High-Automation Vehicles"; it remains in pre-research and drafting-group recruitment, not yet a published standard.
完整战略文档：landscape.md（标准扫描） · proposal-roadmap.md（CSAE / CAICV 预研路线） Full strategy: landscape.md (survey) · proposal-roadmap.md (CSAE / CAICV pre-research path)
ROAM 是开源研究框架（superset），CSAE 团标是其中可标准化的子集（subset）ROAM is the open-source research framework (superset); the CSAE Group Standard is the standardizable subset
2026 年 5 月，CSAE 团标《智能网联汽车 高度自动驾驶车辆运营管理异常事件处置规范》正式启动预研。由中国智能网联汽车产业创新联盟（CAICV）提出，吉林大学、卓驭科技、一汽集团、国家智能网联汽车创新中心（CICV）联合牵头。ROAM 与该团标不等价：ROAM 是开源研究框架，持续国际化演化；CSAE 团标是中国本土的、由起草组共识决定的标准化产出。In May 2026, the CSAE Group Standard Specification for Operations Anomaly Event Management of High-Automation Vehicles entered pre-research, jointly led by Jilin University, ZYT, FAW Group, and the China Intelligent and Connected Vehicles Industry Innovation Center (CICV), proposed by CAICV. ROAM and the CSAE Group Standard are not equivalent: ROAM is an open-source research framework with international scope; the CSAE Group Standard is the Chinese-context, drafting-group-consensus standardization output.
材料状态披露：Material Status Disclosure: ROAM 框架 v1.2 是借助 AI 工具进行结构化整理的初步研究框架，基于公开的国内外标准、法规、行业报告以及作者团队的实践经验。ROAM 中的具体设计（10 种运营模式、三层决策架构、70/25/5 占比、29 子场景、7 维责任矩阵等）均为当前阶段的备选方案，后续可能由起草组进行调整、删减或重新定义。本框架不替代真实运营数据的实证校准、中国法律责任划分的判例支撑、跨企业协作中的真实摩擦点识别，以及监管侧的政治可行性评估。在 CSAE 团标起草过程中，所有条款的最终采纳由起草组专家审议确定。The ROAM v1.2 framework is an initial research framework structured with AI tools, based on publicly available standards, regulations, industry reports, and the author team's practical experience. ROAM's specific designs (10 operating modes, three-layer decision architecture, 70/25/5 split, 29 sub-scenarios, 7-dimension responsibility matrix) are current-stage proposals that may be adjusted, reduced, or redefined by the drafting group. This framework does not replace empirical calibration with real operational data, case-law support for Chinese legal liability allocation, identification of real friction points in cross-enterprise collaboration, or political feasibility assessment from the regulatory side. In the CSAE drafting process, the final adoption of all clauses is determined by expert review of the drafting group.
完整边界说明（中文，给起草组）：07-scope-boundary.md · 立项材料目录：standards/csae-proposal-2026/ Full boundary doc (Chinese, for drafting group): 07-scope-boundary.md · Standard proposal materials: standards/csae-proposal-2026/
Explorer 独立部署正在迁移，当前稳定入口为 GitHub 事件库、项目文档和 Ask ROAM 提问页The standalone Explorer deployment is being migrated. Stable entry points are the GitHub incident database, project documents, and Ask ROAM Q&A page.
交互式 Explorer 恢复前，公开数据以 GitHub YAML / Markdown 为准，页面会持续保留稳定链接 Before the interactive Explorer is restored, public data is maintained as GitHub YAML / Markdown with stable links on this page.
ROAM 是开源基线，各方可在此基础上构建定制化平台An open baseline — build your customized platform on top
Waymo、百度 Apollo Go、Pony.ai、文远知行、滴滴等 Robotaxi 服务运营方Waymo, Baidu Apollo Go, Pony.ai, WeRide, DiDi and other robotaxi operators
Tesla、蔚来、小鹏、极氪、SAIC、一汽等 —— 从制造商到 L4 服务提供者的转型Tesla, NIO, XPeng, Zeekr, SAIC, FAW — transitioning from manufacturer to L4 service provider
Hertz、神州租车等向 L4 车队服务转型的租赁公司Hertz, leasing companies transitioning to L4 fleet service
公交公司、园区运营方、轻舟/宇通等 L4 巴士服务方Bus companies, park operators, L4 bus providers like QCraft/Yutong
美团、京东、Nuro 无人配送；矿区/港口/机场自动驾驶Meituan, JD, Nuro delivery; mine/port/airport autonomous operations
工信部、公安部、地方交通管理部门、保险公司、高校和科研机构Government agencies, transport authorities, insurance companies, universities and research institutions
从单一 Robotaxi 远程运维到覆盖 10 种运营模式的数据驱动风险管理生态From single-mode Robotaxi to a data-driven risk management ecosystem across 10 operating models
当 L4 自动驾驶从 Robotaxi 走向 OEM 订阅、个人共享、公共交通、物流配送等多种模式，每家运营商/OEM 在每个城市都建独立的远程运维团队就像每个车主自己养一支修车队一样不经济。未来远程运营将演化为跨品牌、跨模式的共享基础设施。As L4 expands from Robotaxi to OEM subscriptions, personal sharing, public transit, and logistics, each operator/OEM maintaining independent remote operations teams in every city becomes as uneconomical as every car owner maintaining a private repair crew. Future remote operations will evolve into cross-brand, cross-mode shared infrastructure.
第三方平台为多家运营商提供7×24小时远程监控与异常处置，按车·月收费。类似AAA道路救援——多品牌车辆共享同一套救援基础设施。Third-party platforms provide 24/7 remote monitoring and anomaly resolution for multiple operators. Like AAA roadside assistance — multi-brand vehicles sharing common rescue infrastructure.
远程处置无法解决时，保险公司承保、运维平台执行：派线下团队到场处理拖车、乘客安置、医疗转运。"保险+运维"协同模式。When remote intervention fails, insurers underwrite while operations platforms execute: on-site towing, passenger relocation, medical transfer. An "insurance + operations" synergy model.
基于事件数据库和自然驾驶基线，为不同运营商、ADS系统、城市环境计算差异化事故率。区分"可预防事故"与"不可预防事故"，实现保费精算。Using incident data and naturalistic driving baselines to calculate differentiated accident rates across operators, ADS systems, and cities. Distinguishing preventable vs. unpreventable accidents for actuarial pricing.
远程运维体系标准化，GB安全档案落地，ROAM积累结构化事件数据Remote ops standardization, GB safety case implementation, ROAM accumulates structured data
第三方远程运维服务商出现，保险公司推出Robotaxi专项产品，差异化定价成为可能Third-party service providers emerge, insurers launch robotaxi products, differentiated pricing becomes feasible
"运营商+运维服务商+保险公司"三方协同生态形成，数据驱动的风险管理成为行业常态"Operator + service provider + insurer" tri-party ecosystem forms, data-driven risk management becomes industry norm
详细分析见文献综述第7章 中文版 Full analysis in Literature Review Section 7 English version
四个模块与 17 项国内外标准 / 法规 / 调查报告形成系统性对齐（v1.2 新增加州 DMV、YD/T 4778-4783、Markey RAOS 三组）Four modules systematically aligned with 17 standards / regulations / reports (v1.2 adds California DMV, YD/T 4778-4783, Markey RAOS)
| 标准 / 法规Standard / Regulation | 与 ROAM 的关联Relevance |
|---|---|
| 加州 DMV Article 3.7（Express Terms 2026）California DMV Article 3.7 (Express Terms 2026) v1.2 | 全球首部远程运营专门法规：14 项安全档案 + Remote Driver/Assistant 二分 + § 227.46 分级限制 + § 227.42(i) FRIP 模板First dedicated remote-ops regulation: 14 safety case elements + Remote Driver/Assistant dichotomy + § 227.46 graduated restrictions + § 227.42(i) FRIP template |
| YD/T 4778-4783 v1.2 | 中国 5G 远程遥控驾驶信息交互系列行业标准（6 部，覆盖系统架构、数据采集、信息编码、网络协议、安全要求、测试方法）CN 5G tele-driving information interaction industry standards (6 parts: system architecture, data acquisition, info encoding, network protocol, security, test methods) |
| 美国国会 Markey RAOS 报告（2026-03）US Congress Markey RAOS Report (2026-03) v1.1 | 首份针对 RAO 的国会调查，七家头部 AV 公司均拒绝披露干预频率，从立法层面印证 ROAM 透明度命题First Congressional investigation into RAOs; all seven major AV companies refused to disclose intervention frequency, validating ROAM's transparency thesis |
| GB《自动驾驶系统安全要求》 | 附录 C.2 L4 远程协助要求 · 附录 D 安全档案结构C.2: L4 remote assistance · D: Safety Case structure |
| UN GTR on ADS (WP.29) | "称职且谨慎的人类驾驶员"全球安全基线"Competent and careful human driver" global safety baseline |
| ISO 34502 / ISO 21448 | 场景分类对齐 · SOTIF 未知不安全场景识别Scenario taxonomy · SOTIF unknown unsafe scenarios |
| ISO 22737 (LSAD) | 含"无人驾驶运营调度员"角色定义Includes "driverless operation dispatcher" role |
| 德国 StVFernLV (2025-12)German StVFernLV (2025-12) | 欧洲首个远程驾驶法律框架Europe's first teleoperated driving legal framework |
| NHTSA SGO | ADS 事故 30 秒报告要求ADS crash reporting within 30 seconds |
完整 17 项映射见 related-standards.md · 加州 DMV 全文映射：california-article-3-7-mapping.md · 文献综述：中文 / English Full 17-item mapping: related-standards.md · DMV full mapping: california-article-3-7-mapping.md · Literature review: literature-review.md
ROAM 是社区共建项目。提交事件记录、完善分类体系、改进架构文档，或贡献基线数据。ROAM is community-driven. Submit incidents, improve taxonomy, enhance architecture, or contribute data.
贡献指南How to Contribute