文 | 斯马克开源
两个数字
2026 年 2 月的上海。
一台正在干活的机器人忽然倾覆，机身摄像头和几组精密传感器摔变形。
一个月后5976.87 元到账，人保财险确认，这是国内具身智能行业第一笔保险理赔。
湖北襄阳，在一家科技馆。
一个 5 岁女孩围在圈外看机器人表演，被一记回旋踢踢掉 4 颗牙，面部缝了 3 针。家属把已发生的医疗费、交通费、护理误工费，加上后续疤痕与牙齿修复的费用，测算为 42 万余元，后来把诉求降到 15 万元。
一年多过去了，还没有结论。
让人沉默的不是赔偿谈不下来。是家属反复索要的东西 —— 操作资质、培训记录、运行日志、调查报告 —— 截至公开报道，涉事各方都没拿出来。
机器自己摔了，一个月内就把钱赔付了。
不是赔不起，是我们还欠缺一套规矩。
第一版草稿已经有人替我们写好了，作者就是保险公司。
谁肯为自己的错误标价
保险公司要是判断错了，是真会赔钱。
保险的商业模式是"用概率换现金"，收一块钱保费，就得押上可能几十块的赔付。所以他们必须比任何人都更想知道会不会出事、频率是多少、一次要赔多少。
2025 年 9 月到 2026 年 1 月之间，人保、太保、平安先后把机器人写进保单，那是这个行业第一次被一个外部机构从头到脚摸了一遍。
人保财险浙江分公司在数贸会上首发具身智能机器人专属保险。
没有历史赔付数据，怎么办？
参考工业机器人等相近品类的数据打底，设一个浮动费率区间，再按使用性质校准 —— 同样 100 万元保额的第三者责任险，保费从 200 元到 1000 元一台不等。
拆解一份保价单
头部险企的口径不完全一样，但通过综合对比，能拼出一整套"机器人风险画像"的方法论。
保险公司认为一台机器人的风险由什么决定。
太保产险公开过自己的定价维度，机器人价值（硬件成本）、应用场景风险等级 —— 明确举了舞台表演、工业巡检、商场迎宾三类——再加上维修成本，以及预估的行业与区域风险水平。机型的技术成熟度、核心部件价值也纳入系数。
平安产险引入"动态风险因子调整机制"，把运行时长、作业环境等级、历史安全数据放进定价公式，做成基础保障 + 场景浮动的方案。
人保财险是把一台机器人拆成大脑、小脑、身体三层分别打分：大脑是算法决策，小脑是运动控制，身体是硬件关节，再把运行日志和故障记录接进来动态调整费率。大脑升级快，费率就高一些。身体不会变，费率相对自然低一些。
拼起来我们会得到一张很完整的因果公式：
基础保费 × 场景因子 × 机型因子 × 历史因子
＝ 该交的钱
把这公式翻过来看。同时也是一张产品改进地图。
- 机器贵，维修成本高 → 保费高 → 应该把易损件做成模块化快换。
- 场景人流密集 → 保费高 → 应该做物理隔离和电子围栏。
- 模型升得勤 → 保费高 → 应该把算法版本和故障记录做成可审计的日志。
- 从不交运行数据 → 保费高 → 应该主动上交。
保险公司用保费价格，把这个行业最该做的事，一件一件标上了价码。
以前是没人做。
现在的情况则是国标只能规定下限——告诉你不能低于什么，但不能告诉你现在是多少。
投资人只看增长，不关心你这台会不会伤人。
客户只会看演示——看见的是三分钟的成功，看不见日常使用中的失败率。
只有保费是一个把"会不会出事"直接换算成钱的指标。
保费即评级。
与其听厂商讲一小时，不如让他出示一张保单。
车险能定价，机器人为什么这么难
既然保费这么有用，为什么现在的机器人保险还这么粗糙？
人形机器人保险的核心矛盾，是高风险、快迭代、缺数据，而保险这门生意天生需要的大数据、稳定、可量化。
快迭代 vs 稳定
汽车是相对静态的产品。
买一台车开三年还是那台车。
就算有 OTA，大多也改在座舱和车机，偏辅助性质。保险公司可以给一辆 2023 款的车定价，因为 2024 款基本还是那台车。
而机器人模型的每次升级都可能改动决策逻辑——今天看到障碍物会急停，下个月更新完，可能选择绕过去。
对于他们来说这就是忒修斯之船悖论，因为保险标的在保险期间内会变成另一个东西。
这一点在保险史上似乎没有先例。
一家保险公司给一台机器人签了一年保单，结果年中它经历了四次算法大版本更新——结果这家公司为四个不同的风险体承保，却只收了一份保费。
拿着去年的天气预报预测明天的天气。
技术路线的分歧让情况更加复杂。
现在主流的几条路线 —— 端到端 VLA、世界模型、类脑架构 —— 对保险行业来说就是麻烦。
拿 VLA 机器人的出险数据去给一台世界模型机器人定价，好比拿小轿车的事故率去给重卡定价。
缺数据 vs 大数据
精算靠的是大数法则，要有足够多的样本，才能让概率收敛。
车险有几十年、上亿台车的赔付记录。
2026 年中国具身智能机器人商业化刚刚起步，历史赔付样本几乎是空白。
不同型号、不同算法系统、不同使用场景、不同操作人员水平，都会导致风险差异巨大。
商场迎宾、舞台表演、教育研学、文旅互动、企业展厅接待 —— 看着都是"商业应用"，实际风险完全不同。
并且险企难以实时获取振动、电流、温度等数据，主动预警与动态风控能力不足。
保险公司想检查的那些物理量 —— 关节电流、电机温度、减速器振动 —— 本是判断一台机器能否继续使用的，最有价值的信号。
而这些数据今天大部分捂在厂商手里。
高风险 vs 可量化
"高风险"不只是出事概率高，更是一旦出事，损失不可预测。
车险的损失结构很清楚，撞了就是撞了，修车多少钱有配件价格库，人伤有伤残等级标准。
整个社会的定损体系已经跑了三十年。
机器人的损失结构里可能有算法偏差、系统漏洞、以及网络劫持导致的物理损害，这些在传统企业财产险里根本没有对应科目。
机器人的属性和保险的前提，很难兼容。
而这正是我们接下来要看到的种种怪象的源头。
摔了以后，谁说了算
人保财险公开讲过定损的做法：目前市场缺乏针对具身智能机器人的第三方评估机构和维修网络，所以"我方与主机厂签署合作协议，由主机厂进行损失确定，我方按合同约定核定损失金额"。
你的机器人摔了，赔多少钱，由造它的那家公司说了算。
在汽车行业这是不可想象的。
没有任何一家车险公司会接受由比亚迪来定损比亚迪，由特斯拉来定损特斯拉。
汽车行业前后花了三十年，才建起独立的公估体系、公开可查的配件价格库和社会化的维修网络。
机器人行业目前一样都没有。
于是长三角某租赁平台首批 470 余台机器人，如果因碰撞或传动装置断裂毁损，每台最高可获得近 40 万元的维修费用赔偿。
而实际第一笔赔出去的钱，是 5976.87 元。
一台机器人摔了，摄像头坏了，值多少钱？
关节松了，值多少钱？
某个模组烧了，值多少钱？
只有造它的那家公司清楚。
公开可查的几张单子里，本体损失保额有 20 万、40 万两档，第三者责任有 10 万、30 万、50 万三档。
记者问到擎天租的三者险，得到的答案是——"尚未发生此类赔付案例"。
也就是有关于机器人的人身伤害，中国至今是零赔付样本。
零赔付有两种可能性。
一种是机器人真的很安全。
这是好消息，但无法验证息——因为还没建立起让事故被强行记录下来的机制。
另一种是出过事，只是没走保险。
没有赢家的现场
2026 年 3 月，山东临沂一家商场
四台人形机器人在台上表演武术。
工作人员拉了警戒线，也反复提醒别靠近。但总有好奇的家长把孩子往前推。
一个小男孩翻过警戒线，伸手去摸机器人，正好赶上其中一台做连贯踢腿动作，机械腿结结实实碰到了孩子。
相关公司第一时间送孩子去医院检查，最后全额承担了医药费和营养费。
2026 年 5 月前后，武汉一所学校
两台机器人去做活动表演。
筹备阶段，工作人员手滑误触了操作屏，机器人直接进入阻尼模式，起身瞬间重心没稳住，机身摄像头当场摔裂。
这次属于操作失误，维修费由具身智能公司自己承担。
2025 年 12 月，杭州一家餐饮店开业
请了机器人撑场面，现场铺了厚厚的红毯。
机器人走上去重心偏移，一头栽倒，砸坏了店门口的饮料机。
主办方说是操作不规范，租赁方说我提前告诉过你地面要平整。最后各回各家，损失各担一半。
湖南，一台"裸租"的机器人
客户只租机器、不带操作员。
还回来以后，机械臂关节有点卡顿，租赁商按返厂维修价从押金里扣了一笔。
客户不乐意，说这机器租之前就有毛病，告上法庭。
租赁商手里证据不全，打官司耗时耗力，最后押金全额退了回去。
这些案例有一个共同点：
甲乙双方都没有清晰的责任判定。
操作失误的，自认倒霉。
场地问题的，双方私了。
设备老化的，押金就是唯一的制衡手段。
机器人行业至今在用最原始的机制处理纠纷 —— 谁手里扣着钱谁说了算。而一旦走司法程序，又因为"证据不全"而退回原点。
2025 年 1 月，澳大利亚昆士兰，一台工业机械臂在维护后重新启动，当场把一名工人撞倒在地，三根肋骨骨折。
调查发现，厂里根本没有设备隔离程序，工人进入作业区前机械臂没有被切到安全状态。公司被罚 12.5 万澳元。
责任其实很清楚——工厂没有安全程序。之所以能被查清楚，是因为有独立的调查机制。
国内案例里，我们要么靠私了，要么靠押金，要么靠一方的自觉。
没有任何一起，是靠一份双方都认可的记录。
一次事故，四方主体，五张保单
- 硬件制造商： 产品责任，确保机器人物理结构的安全性
- 算法提供商： 技术瑕疵责任，保障算法的可靠性与可解释性
- 运营使用者： 管理责任，确保机器人在合规场景下正确使用
- 系统集成商： 配置责任——围栏、任务设定、感知参数的改动
一次事故可能同时涉及这四方。而理论上，会有五张保单同时响应。
- 工伤保险
覆盖劳动者，不覆盖制造商。赔付后会向责任方代位追偿——所以只是把问题往后推了一步。
- 一般责任险（CGL）
在美国，从 2026 年 1 月开始，可能已经被 AI 除外批单剔掉了。
- 产品责任险
只在把 AI 软件认定为"产品"的司法辖区才成立。而认定各地不一 —— 欧盟新版《产品责任指令》明确把软件算作产品，其他地方还在争论。
- 网络安全险
保的是劫持和数据泄露，不保物理伤害。
技术 E&O。
通常要求"人类在旁边监督"才赔。而如果一个系统宣称全自主，可能直接失效。
五张保单，可能同时响应，也可能一张都不清晰地响应。
每一张都能找到一个理由说"不归我管"。
也就是所谓的「静默 AI」（Silent AI）。
"保单里既没说保，也没说不保"。
保单是为一个没有 AI 的世界写的。等 AI 造成的损失来了，保险公司和投保人都只能去翻那些从来没打算处理这种情况的条款。
2017 到 2020 年涌现出一批「静默网络风险」（Silent Cyber）危机，当时大量网络损失砸在从来没有设计过要覆盖它们的保单上 —— 一般责任险、财产险、海上货运险。
行业为此耗费了三年、依靠两份劳合社市场公告和一轮全市场整改勉强渡过了这个阶段。
什么是盲保？
传统财产险、设备险需要明确设备的存放和使用地址，而具身智能机器人保险不约定具体使用地址，只要在中华人民共和国境内使用均可投保。
对照车险想想就明白问题在哪。
车险要问常年停放地、行驶区域、年行驶里程。因为一辆车的风险不在"车是卡罗拉还是特斯拉"，在于车辆每天跑多远、停在哪、是谁在开。
保险公司从来不给"一台车"定价，只给一台车 ＋ 一种用法的范畴内定价。
机器人同理。
风险不在"是什么型号"，而在它在哪、在干什么、跟多少人接触。
而眼下这三个变量，全是失控的状态。
- 租赁模式下的机器人天天换场地，今天是商场中庭，明天是工厂车间
- 具身智能租赁场景里人机比接近一比一 —— 一台机器配一个人，"人"这个最大的不确定源，始终在场
- 不同应用场景下保费差别最高可达三成以上 —— 承保人自己对场景风险的分层能力还很粗糙。
为什么不"不约定使用地址"？
因为承保人还没想好怎么定价，只好先不定价。
盲保不可怕，可怕的是盲保的时候，被保对象正在大规模进入公共空间。
截至 2026 年 10 月，国内尚没有任何一条直接针对公共场景人形机器人表演的全国性强制性安全标准。
安全距离、物理隔离、操作员资质、急停机制、事故上报 —— 最容易出事的环节，全部处于无强制约束状态。
- 临沂，是安全距离和物理隔离的问题
- 武汉，是操作员资质和误触防护的问题
- 襄阳，是安全距离加操作资质的问题
- 杭州，是场地评估的问题。
当前具身智能产业的突出问题是躯体太强、大脑太弱，实验室里表现不错，进入真实世界后面对突发情况的应对能力仍然不足。
力气越大，能力越强，出事的后果越重。
而当下承保人能拿到的信息只有：是什么型号，值多少钱。
2025 年 9 月，太保产险宁波分公司推出"机智保"。
支持按天、按周、按月投保，突破传统年度保单限制。以"风险拟人化"理念把本体损失、第三者人伤与财产损失整合进一张保单，贯通生产、销售、租赁、使用全链条。
- 人保财险
把保障一路延伸到研发端——机器人还在概念验证和中试阶段，失败造成的费用损失也能保，这个延伸很聪明。
2026 年 1 月，平安产险
联合上海电气融资租赁落地全国首张具身智能机器人"保险 ＋ 租赁"保单，保费直接摊进租金。
企业不用一次性掏一大笔钱，保障范围还包括产品质量责任和信息泄露责任。
真金白银的政策也有：
- 北京经开区：对企业投保按实际保费 50% 补贴，每年最高 100 万元
- 宁波：对保费费率不超过 3% 的企业，按实际投保年度保费 80% 补贴，同一企业每年最高 200 万元。
数据有别的问题
对比美国的政策，"中国赢了"吗？
不能这么说。
因为美国承保人的选择有非常扎实的依据。
他们确实没有可以定价的东西
精算的前提是损失三角——要有足够的历史赔付记录，才能画出损失随时间的分布。
而生成式 AI 商业化才三年左右，已报案的样本里绝大多数还是版权和知识产权纠纷，不是人身伤害。
保单真正要覆盖的那类损失（人身伤害、财产损害），恰恰是数据空白。
他们自己也没统一意见
- AIG 在爱达荷和伊利诺伊备了案，却明确告诉《金融时报》"目前无实施计划"。
- S&P Global 统计：41 家 P&C 集团至少有一家子公司备案除外生成式 AI，另有 20 家专门备案"延迟"。
- Marsh McLennan CEO John Doyle 在 10 月 15 日的财报会上公开警告：宽泛的 AI 除外会造成"真正的困惑与不确定性"，主张更细致地看待风险如何聚合。
龙头经纪公司的 CEO 都在公开唱反调 —— 这不是一个想清楚了的市场环境。
最重要的是必须划清一条边界
他们定义的对象是"生成式 AI"，未明示覆盖只做分类或评分的预测型模型。
而这个边界，随着每一次模型更新都在移动 —— 厂商不断给原本是预测型的工具加上生成式功能。
所以不能说"美国把机器人责任全部撇除了"。
准确的说法是：美国正在把"能生成内容的 AI"造成的损害，从标准保单里往外搬。
但全美州备案审查发现，没有任何一款 admitted 的独立产品，能接住这些批单剔掉的那部分保障。
也就是说到底，一家美国工厂买了机器人，要是出了事，保险公司说这是 AI 引起的，不赔。
那钱谁出？
工厂出。或者，工厂去起诉机器人厂商。
也就是说，除外条款不是在消灭风险，是在把风险从可定价的日常成本变成"不可预测的或有负债"，和美容整形行业没什么区别。
对一家上市公司来说，这两者的差别巨大。
中国在补课
第一，示范条款
2026 年 3 月，中国保险行业协会明确表示将逐步推进人形机器人保险示范条款编制，聚焦核心风险场景，统一保障边界与责任界定，覆盖硬件本体损失、算法故障、网络安全及第三方人身财产损害，并规范保费定价逻辑与理赔流程。
示范条款一旦落地，意味着不同公司的保单第一次可以横向比较。
而横向可比，是"保费即评级"能够成立的前提。
在此之前，各家条款自定、限额协商约定——据《每日经济新闻》查阅备案产品目录，主流产品的责任限额与免赔额均采用"协商约定"模式。
拿到两份报价也无法比较，因为保的东西可能根本不一样。
第二，保障在往颗粒度更细的地方发展
2026 年 10 月，太保产险落地国内首单"灵巧手"具身智能关键零部件保险，保障设备存放及使用中的意外事故、自然灾害、设计安装与材料缺陷、人员误操作带来的损失，同时覆盖第三方人身伤害与财产损失。
把保险做到零件级别，看起来很繁琐，实际上是承保人开始有信心对更小的风险单元定价了。
第三，保障在往产业链更前端发展
人保财险把保险做到了研发端——概念验证综合费用损失保险、中试验证综合费用损失保险。
具身智能企业最怕的不是机器摔了，是投了几百万做完中试发现路线走不通。
说明承保方已经理解了这个行业的真实风险结构。
第四，再保险在跟进
中再产险早在 2022 年就落地了国内首单移动机器人保险，2026 年进一步把人形机器人、四足机器人、工业机械臂全部纳入保障体系。
再保险是保险公司的保险公司，风险池被认为是可以分散的。
第五，从单台投保转向场景批量承保
传统做法是单台零散投保。
而现在，人保为长三角某租赁平台首批 470 余台机器人一次性承保，平安通过租赁场景批量承保。
可以依托场景数据，破解数据壁垒与需求碎片化。
批量承保是数据获取的工程方案，保险公司并不是在做慈善。
车架号
车险之所以能定价，第一个前提其实是每辆车有一个唯一的车架号。
没有 VIN，就没有出险记录可以挂载，没有二手车残值。
没有 UBI，没有按车型统计的事故率。
整个车险大厦的地基，是一串十七位的字符。
2026 年 5 月 22 日，《人形机器人 全生命周期管理规范》标准发布。
每台人形机器人获得一个 29 位唯一、不可更改、贯穿全生命周期的身份编码。
配套的人形机器人全生命周期管理服务平台，在工信部科技司指导下搭建，覆盖研发、生产、准入、销售、使用、维护、报废、回收全链条。 比如宇树科技所有机器的"一机一码"已全面接入平台。
编码规则制定的难点在兼容与共识——各头部厂商已经建了自己的内部编码体系，标准要考虑兼容性和可扩展性，最后才定下四段式结构。
更换关键零部件时，服务商必须获得管理方授权，并重绑编码、重新测试，回收前应彻底清除存储数据、解除所有绑定。
"更换关键零部件必须重绑编码并重新测试"——如果执行到位，会解决"裸租纠纷"里最核心的问题：机器在租赁期间到底被谁动过。
从纸面标准变成行业准则，还有很长的路。
让利益相关方"听得懂、用得上"。
现在该做什么？
阿西莫夫写了六十年机器人，反复探讨的其实只有一个问题：
出事了，算谁的。
机器人三定律不是写给机器人读的，本质上是一份责任分配协议。
七十年后我们真的造出了人形机器，能轻松一脚踢掉孩子好几颗牙齿。
现在有人已经开始率先替我们起草协议了，他们正是伟大的保险公司。
但他们做的再好，还是取决于机器人行业愿不愿意交出黑盒子里的运行日志。
参考信源
- 中共上海市委金融委员会办公室《一台机器人意外摔倒获赔 5976 元，全国首例机器人保险理赔在沪落地》
- 人民网财经《机器人租赁火了 保险如何兜住商业化风险?》
- 经济参考报：具身智能机器人综合保险双轨体系、不约定使用地址、人机比 1:1、三档限额及"尚未发生赔付案例"
- 《财经》客户端《机器人走进生活，风险谁来"抗"?》
- 中国金融新闻网《机器人保险距离规模化落地还有多远?》《人形机器人赛场竞速 保险保障同步"陪跑"》
- 中国经营报《从资本托举到风险兜底 保险深度入局机器人产业》
- 证券时报网《人形机器人也能上保险了，面临哪些痛点和难点?》《人形机器人"上保险" 护航具身智能走出实验室》
- 新华网《具身机器人也有自己的保单 险企竞逐机器人保险业务》
- 经济日报《给人形机器人上保险》《平安产险携手上海电气共建智能制造风险保障新生态》
- 中国经济网《人形机器人上保险 风险轮廓还需摸得清》《给人形机器人"上保险"，还需迈过几道坎?》
- 中证网 / 中国证券报《给人形机器人上保险!险企布局新赛道》
- 新华财经：人保财险浙江分公司浮动费率区间（三者 100 万保额，保费 200–1000 元/台）
- 新浪财经《人形机器人年产将超 10 万台 险企如何为具身智能机器人"受伤"买单》
- 每日经济新闻《具身智能机器人把我弄伤了，谁来赔?》（责任限额与免赔额"协商约定"模式）
- 京报网 / 央广网 / 国际金融报《中国保险行业协会：将逐步推进人形机器人保险示范条款》
- 常州日报 2026-10-14 A04 金融版（太保产险灵巧手具身智能关键零部件保险首单）
- 中华人民共和国科学技术部《关于加快推动科技保险高质量发展 有力支撑高水平科技自立自强的若干意见》（2026-02，点名具身智能）
- 经济参考报 / 科技日报 / 中国新闻网 / 工人日报 / 中国企业报：《人形机器人 全生命周期管理规范》29 位编码、2.8 万余台赋码、宇树全量接入
- 中国新闻周刊 / 紫牛新闻：襄阳女童被机器人踢伤事件
- 人民日报 / 江津网：广东潮州送餐机器人撞伤幼童案判决
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Smart Open.
Two numbers.
Shanghai, February 2026.
A working robot suddenly fell over and the body cams and several sophisticated sensors fell into disfigurement.
A month later, 59.76.87 came to the books, and PMI confirmed that this was the first insurance settlement in the country ' s smart industry.
Hebei Yiyang, in a technology museum.
A 5-year-old girl, watching a robot show outside the circle, was kicked out of four teeth with three stitches on her face. The family calculated the medical expenses incurred, transportation costs, care costs, plus the costs of subsequent scars and dental repairs, to be over $420,000, and then reduced the claim to $150,000.
More than a year has passed without conclusion.
It's not that you can't talk about compensation. It's what the family has repeatedly asked for -- operational qualifications, training records, operating logs, investigation reports -- and as of the time of the public reporting, none of the parties involved had taken it out.
The machine fell and paid for it in a month.
It's not that we can't afford it. It's that we still lack a set of rules.
The first draft has been written for us, by the insurance company.
Who's going to bid for their mistakes?
If the insurance company is wrong, it will pay.
The business model of insurance is "cash-for-probability" and a dollar premium is charged at a possible cost of dozens. They must therefore be more interested than anyone in knowing whether there will be an accident, how often, and how much to pay once.
Between September 2025 and January 2026, human security and safety put robots into the policy, which was the first time that the industry had been touched by an outside agency.
Zhejiang Sub-Sector of Human Security has launched a special insurance policy for smart robots at trade fairs.
There's no historical data. What do we do?
The reference to the data of comparable products, such as industrial robots, is bottomed up with a floating rate range, calibrated by the nature of the use -- third-party liability for a similar $1 million guarantee, ranging from $200 to $1,000.
Break a policy.
The head is not exactly the same caliber, but a combination of comparisons can spell out a whole set of "robots-risk" methodology.
The insurance company believes that the risk of a robot is determined by what.
Oversure has exposed its pricing dimensions, robotic values (hardware costs), application scenarios risk levels - with three categories of stage performances, industrial inspections, mall hospitality - plus maintenance costs, and estimated industry and regional risk levels. The technological maturity of the machine type and the value of the core components are also factored in.
Safe birth insurance introduces a dynamic risk factor adjustment mechanism, which puts long running hours, operating environment levels, historical safety data into pricing formulas, as a basic security + scenario float.
People ' s financial insurance is to split a robot into a brain, a small brain and a three-tiered body: the brain is algorithmic decision-making, the small brain is motion control, the body is a hardware joint, then the running logs and failure records are attached to the dynamic adjustment rate. The brain upgrades faster, the rates are higher. The body doesn't change, the rate is relatively low.
Together we'll get a complete causal formula:
Basic premium x scenario factor x machine factor x historical factor
= money due
Turn this formula around. It is also a product improvement map.
- The machine is expensive, the maintenance costs are high, the premiums are high, it should be modularized.
- The scene is densely populated and the premium is high and should be made physical isolation and electronic fencing.
- The model works well, the premium is high.
- Never turn in running data → high premium → should go forward.
The insurance company uses the premium price to mark one of the most important things in this industry.
Nobody used to do it.
It is now the case that the national flag can only set a floor — to tell you what it is — but not how much it is.
Investors look at growth and don't care if you're hurt.
Clients only look at demonstrations — they see three minutes of success, and they see no failure rate in daily use.
Only the premium is an indicator that makes money directly to convert.
The premium is the rating.
Instead of listening to the manufacturer for an hour, he should be shown a policy.
Car insurance can be priced. Why are robots so hard?
If the premiums are so useful, why is the robot insurance so rough?
The core contradiction of human robot insurance is the high-risk, fast-temporality, lack of data, and the large data, stable, quantization that the insurance business was born to.
Fast iterative vs stable
Cars are relatively static products.
Buy a car for three years or that car.
Even with OTA, most of it is in the cabin and the engine, which is more auxiliary. The insurance company can price a 2023 car because it's basically the same car.
And every upgrade of the robotic model could change the logic of decision-making — seeing the barriers today would be abruptly halted, updated next month, and possibly bypassing.
For them, this is the Teshus ship paradox, because the subject matter of the insurance becomes something else during the insurance period.
There seems to be no precedent for this in insurance history.
An insurance company signed a one-year policy for a robot, which led to four major updates of algorithms in the course of the year — as a result of which the company covered four different risk groups, but only received one premium.
Take last year's weather forecast and predict tomorrow's weather.
Differences in technical routes further complicate the situation.
Now the main lines -- end-to-end VLA, world model, brain structure -- are trouble for the insurance industry.
Using VLA data to price a world model robot like a car accident rate to price a heavy card.
Missing data vs Big data
The actuarial function is based on the rule of large numbers, with a sufficient number of samples to reduce the probability.
Car insurance has been recorded for decades and hundreds of millions of vehicles.
In 2026, China ' s smart robots were just beginning to commercialize, and historical compensation for sample was almost blank.
Different models, different algorithm systems, different scenarios of use, different levels of operators can lead to significant risk variations.
Commercial hospitality, stage performances, educational research, literary interaction, business fair reception -- it looks like "business applications" with different actual risks.
It is also difficult to obtain data on vibrations, currents, temperature, etc. in real time, with insufficient capacity for active warning and dynamic wind control.
The physical quantities that insurance companies want to check -- joint currents, electric temperature, brake vibrations -- are the most valuable signals to determine whether a machine can continue to be used.
Most of these data are in the hands of manufacturers today.
High risk vs can quantization
The "high risk" is not only a high probability of an accident, but also an unpredictable loss in the event of an accident.
The loss structure of the vehicle insurance was clear, the crash was a hit, the amount of the repair vehicle had a spare parts price bank and the injury had a disability level.
The system of social losses has been running for 30 years.
The loss structure of robots may contain algorithm deviations, system loopholes, and physical damage caused by cyberjackings, which have no counterpart in traditional enterprise property insurance.
The attributes of robots and the prerequisites of insurance are difficult to reconcile.
And that's what we're about to see.
Who's in charge after the fall?
Financial insurance has made it known that the market currently lacks a network of third-party assessment bodies and maintenance of smart robots, so "we have signed a cooperation agreement with the mainframe, which determines the losses to be incurred by the mainframe, and we approve the amounts of the losses as contracted".
Your robot fell, how much did it cost? It's up to the company that built it.
It's inconceivable in the automobile industry.
None of the car insurance companies would accept Biadi to be awarded damages and Tesla to be awarded damages.
It took 30 years for the automobile industry to establish an independent public valuation system, open and accessible spare parts price banks and socialized maintenance networks.
There is no robotic industry.
As a result, the first group of 470 robots on a leased platform in the long triangle could be compensated for maintenance costs of up to $400,000 each, if damaged by a collision or a break-up of a moving device.
And the first amount actually paid out was $5976.87.
A robot fell. The camera broke. How much is it worth?
How much is it worth?
How much is it worth?
Only the company that built it knows.
In the list that is publicly available, there are 200,000, 400,000, and the third is 100,000, 300,000 and 500,000.
The reporter asked about Triple Rent, and the answer was, "There have been no such cases."
In other words, there are physical injuries to robots, and China has so far paid zero compensation for sample.
There are two possibilities for zero compensation.
One is that robot is really safe.
That is good news, but it cannot be verified — because no mechanism has yet been put in place to force the recording of accidents.
The other one was an accident, just no insurance.
There's no winner.
In March 2026, a mall in Shandong Qin
Four human robots perform martial arts on stage.
Staff members pulled the cordon and repeatedly cautioned against approaching. But there are curious parents who push their children forward.
A young boy turned over the perimeter and reached out to the robot to catch up with one of them to do a consistent kick. The mechanical leg was so solid that he touched the child.
The company sent the child to the hospital for examination at the first time and finally fully covered the cost of medicine and nutrition.
Around May 2026, Wuhan School
Two robots for a live show.
In the preparatory phase, the staff members slipped through the operating screen, the robot went straight into the blocker mode, the centre of gravity was unstable at the moment of the rise, and the body cavity collapsed on the spot.
This was an operational error and maintenance costs were borne by smart companies.
In December 2025, a restaurant opened in Hangzhou.
We've got robots on the scene, we've got a thick red carpet.
The robot went to the centre of gravity and fell and destroyed the drink machine in front of the shop.
The sponsor said it was irregular, and the lessor said I had told you in advance that the ground was flat. In the end, we went back home and lost half.
Hunan, a naked rented robot.
Clients only rent machines without operators.
When returned, there was a bit of Cardon at the arm of the machine, and the lessor deducted a deposit from the deposit at the rate of return maintenance.
The client was reluctant to say there was a problem before the machine was rented and brought to court.
The renter had incomplete evidence and the litigation was time-consuming and the deposit was returned in full.
These cases have one thing in common:
Neither party has a clear determination of responsibility.
It's a mistake.
The site is private.
When the equipment is ageing, the deposit is the only countermeasure.
The robot industry has been using the most primitive mechanisms to deal with disputes -- who holds the money and who rules. And once the judicial process is under way, the evidence goes back to where it came from.
In January 2025, in Queensland, Australia, an industrial mechanical arm was restarted after maintenance and a worker was hit on the ground and three ribs were broken.
The investigation found that there was no equipment quarantine procedure at the plant and that the mechanical arm of the worker had not been cut to safety before entering the area of operations. The company was fined $125,000.
The responsibility is clear — there are no safety procedures at the plant. This was made possible by the existence of an independent investigative mechanism.
In domestic cases, we're either on our own or on deposit or on one side's own initiative.
There is no one, but a record that is mutually acceptable.
An accident. Four subjects, five insurance policies.
- Hardware manufacturer: product responsibility to ensure the safety of robotic physical structures
- Algorithmic provider: Liability for technical defects, safeguarding the reliability and interpretability of algorithms
- Operating users: Managing responsibility to ensure that robots are used correctly in compliance settings
- System Integrator: Configure responsibilities — fencing, mission setting, changes in sensor parameters
An accident may involve all four. In theory, there will be five policy responses.
- Work injury insurance
Cover workers, not manufacturers. Compensation is recovered from the subrogation of the responsible party — so it is just a step backwards.
- General liability insurance (CGL)
In the United States, since January 2026, it may have been removed by the AI exception.
- Product liability insurance
Only jurisdictions that have identified AI software as "product" have been established. The new EU Product Liability Directive explicitly classifies software as a product and is still under debate elsewhere.
- Cybersecurity.
Hijacking and data disclosure are protected against physical harm.
Technology E&O.
It's usually called "human supervision." If a system claims full autonomy, it may simply fail.
Five insurance policies may respond simultaneously or one may not respond clearly.
Each one can find a reason to say, "I don't own it."
It is called Silent AI.
"There's no insurance in the policy. There's no insurance."
The policy was written for a world without an AI. When AI incurred losses, both insurance companies and insurers had to go through those provisions that had never been intended to deal with them.
By 2020, a number of Silent Cyber crises had emerged, with massive network losses. On policies that have never been designed to cover them - general liability insurance, property insurance, maritime freight insurance.
The industry has spent three years on this and has gone through this stage with two LAS market announcements and a round of market-wide restructuring.
What's blind?
Traditional property and equipment insurance requires a clear location and address for the use of equipment, while the smart robot insurance does not agree on a specific use address, which can be insured for use in the territory of the People ' s Republic of China.
Think about the risk of a car and see what the problem is.
Car insurance requires constant parking, areas of travel and miles per year. Because the risk of a car is not "Carola or Tesla" but how far, where and who is driving every day.
The insurance company never priced "one car" but only one car plus a one-use frame.
Robots are the same.
The risk is not "what type," but where it is, what it is doing, how many people are in contact.
And now these three variables are out of control.
- The robots under the lease model are changing places every day. Today is the middle of the mall. Tomorrow is the factory.
- Smart lease scenario with a human machine closer to one -- a machine with one person, the biggest source of uncertainty, always present.
- The difference in premiums between different applications can be up to 30% - the insurer ' s own layers of risk to the scene are rough.
Why don't you "not agree to use the address"?
Because the insurer didn't figure out how to price it, so we didn't.
Blind coverage is not scary, it is when the target is entering public space on a large scale.
As of October 2026, there is no national mandatory safety standard that directly targets human robotic performances on public scenes.
Safe distance, physical isolation, operator qualifications, stop mechanisms, accident reporting - the most vulnerable link, all in a non-coercive state.
- Pelicans. Safe distance and physical isolation.
- Wuhan, it's the operator's qualifications and fault protection.
- Jianyang, it's a matter of safe distance and operational qualification.
- Hangzhou, the site assessment.
The salient problem in the current smart industry is that the body is too strong, the brain is too weak, the laboratory is doing well and the ability to respond to emergencies behind the real world remains inadequate.
The greater the strength, the greater the capacity, the greater the consequences.
And the only information the underwriter can get is what type it is, how much it's worth.
In September of 2025, Tai Peng Ning Po introduced "The Smart Man".
Support is provided for daily, weekly and monthly insurance coverage beyond the traditional annual insurance policy limits. The concept of "risk-to-humanized" integrates the underlying loss, third-party injury and loss of property into a single policy, linking production, sale, lease and use of the whole chain.
- People's insurance.
Extending safeguards to the research and development end — robots are still in the conceptual validation and mid-test stages, and the cost of failure can be saved, and this extension is smart.
Safe motherhood insurance, January 2026
The country's first insurance policy for smart robots in the United Shanghai Electrical Finance Leasing, with the premium being allocated directly to the rent.
Enterprises do not have to pay a large sum of money at once, and the scope of safeguards also includes product quality and information disclosure responsibilities.
The real-gold policy is also:
- Beijing Economic Zone: insurance of enterprises on the basis of actual premiums 50 per cent of the subsidy, up to RMB 1 million per year
- Ningbo: 80 per cent of the annual premium for an enterprise with a premium rate not exceeding 3 per cent is paid, up to 2 million yuan per year for the same enterprise.
There's something wrong with the data.
Compared to American policy, "China wins"?
You can't say that.
Because there is a very solid basis for the choice of American insurers.
They don't have anything to price.
The actuarial premise is the loss triangle, which requires adequate historical compensation records to produce a distribution of losses over time.
The majority of the reported sample were copyright and intellectual property disputes, not personal injury.
The type of loss (personal injury, damage to property) that the policy is really intended to cover is precisely a data gap.
They don't agree.
- AIG filed a case in Idaho and Illinois, but explicitly told the Financial Times that there was no implementation plan.
-S&P Global Statistics: 41 P&C groups have at least one subsidiary to file except generation of AI and 20 dedicated "delays".
- Marsh McLennan CEO John Doyle publicly warned in a press conference on October 15th that broad AI except would create "real confusion and uncertainty" and advocate a more detailed view of how risks converge.
The CEOs of the Dragon Brokers are all singin' out loud -- it's not a clear market environment.
The most important thing is to draw a border.
They define "generic AI" and do not explicitly cover predictive models that are classified or rated only.
And this boundary, as each model is updated, is moving -- the manufacturers are adding a generation function to what is originally a predictive tool.
So we can't say, "The United States has left all the robots behind."
To be precise, the United States is moving the damage caused by "AI" that generates content from the standard insurance policy.
However, the U.S.C. filing review found that there was no single independent product that could capture the part of the security that had been removed from the list.
In the end, an American factory buys robots, and if anything happens, the insurance company says it's AI, and it doesn't pay.
Then who pays?
The factory's out. Or the factory goes to sue the robot manufacturer.
In other words, the exclusion clause is not eliminating risks, it is transforming them from priced daily costs to "unforeseeable contingent liabilities" and is no different from the cosmetics industry.
For a listed company, the difference between the two is significant.
China's making amends.
First, model clauses
In March 2026, the China Insurance Industry Association made it clear that it would gradually advance the development of model provisions for human robotics insurance, focus on core risk scenarios, harmonize border and liability definitions, cover the loss of hardware capital, algorithmic malfunctions, network security and damage to third-party personal property, and regulate the pricing logic and compensation process.
Once the model clause had landed, it meant that policies of different companies could be compared horizontally for the first time.
And horizontal comparability is a prerequisite for a premium or rating.
Prior to this, individual terms and limits were negotiated - the catalogue of registered products was consulted in the Daily Economic News, and the liability limits and exemptions for mainstream products were based on the "consultative agreement" model.
It was not possible to compare the two offers, since the security might not be the same at all.
Second, we're going to be able to move to a more nuanced area.
In October 2026, the country's first smart, smart, critical spares insurance policy in the risk-taking home country, guaranteed loss from accidents, natural disasters, design installation and material defects, human error, and covered third-party personal injury and loss of property.
Making insurance at the level of the spare parts appears cumbersome and, in fact, the insurers are beginning to have confidence in pricing smaller risk units.
Third, to ensure a more advanced development in the industrial chain
Financial insurance covered the research and development end-of-the-art insurance against the loss of the combined cost of proof of concept and the loss of the combined cost of the medium test certificate.
It was not the machine that was the most afraid of smart firms that fell, but the millions that had finished the test and found that the route was not working.
It indicates that the insurers have understood the real risk structure of the industry.
Fourth, reinsurance is on track.
In 2022, China's first mobile robot insurance was set in place, and in 2026 it was further integrated into the safeguards system.
Reinsurance is an insurance company and the risk pool is considered to be dispersed.
Fifth, move from single-stage coverage to mass coverage.
The traditional practice is to have separate and sporadic insurance coverage.
And now, one of the first 470 robots on a lease platform in the Long Triangle is insured by a lot of the lease.
Data barriers and demand fragmentation can be broken depending on scene data.
Batch insurance is the engineering programme for data acquisition and insurance companies are not charitable.
Garage number
The first condition for the pricing of vehicle insurance is that each vehicle has a single shelf number.
Without VIN, there is no risk record to mount and no used car residual value.
No UBI, no car-type accident rate.
The foundation of the entire vehicle safety building is a series of 17-bit characters.
On May 22, 2026, a standard publish of the Code of Practice for the Management of the Full Life Cycle of Human Robots.
Each human robot gets a 29-bit unique, immutable, life-cycle identity code.
An accompanying platform for the life cycle management of human robots was established under the guidance of the Science and Technology Department of the Ministry of Industry and Communications, covering the entire chain of research and development, production, access, sale, use, maintenance, end-of-life and recycling. For example, the one-size-one machine of all of the Ukri technology machines is fully connected to the platform.
Coding rules are difficult to develop in terms of compatibility and consensus — heads of manufacturers have already built their own internal coding systems, standards that take into account compatibility and scalability before finalizing the four-part structure.
In the replacement of critical parts, the service provider must obtain management authorization and recodify, retest and remove the stored data and untie all before recovery.
"Replacement of critical parts must be recoded and retested" - if implemented, it would solve the central problem in the "naked lease dispute": who moved the machine during the lease.
There is still a long way to go from paper standards to industry guidelines.
Let stakeholders "understand and use."
Now what?
Asymov wrote 60 years of robotics, and there's only one problem he's been talking about:
Something's wrong.
The three laws of robots are not written to robots. They are essentially a duty-sharing agreement.
Seventy years later, we actually built a human machine that could easily kick several teeth out of the kid.
Now someone has begun to draft an agreement for us. They are the great insurance companies.
But the best they can do is the robot industry's willingness to hand over the operating logs in the black box.
Reference Sources
- The CCP Shanghai Finance Committee Office, "A robot fell and lost 5976 dollars, and the country's first case of robot insurance fell on the ground."
- The People's Net. How does the insurance cover the risks of commercialization?
- Economic Reference Post: Integrated Security System for Smart Robots, non-agreed address, 1:1, 3-tier limit and "no outstanding claims"
- The Money client. "The Robot goes into life. Who's at risk?"
- China Financial News, How far is the robot's safety from mass?
- The Chinese newspaper, "From capital endowments to the bottom of risk, insurance depth into the robot industry."
- Security Times, "Man-shaped robots can be insured, what are the pains and difficulties?" - "Human-shaped robots" - "Insurance" to escort body smarts out of the lab.
- Xinhuanet, the robots have their own insurance policy.
- The daily Economics, " Insurance for human robots ", " Safe Production Risks Together to Build a New Ecology for Risks of Intellectual Manufacturing in Shanghai ".
- China's economic network, "Insurance on human robots, risk contours need to be identified." - "Insurance on human robots."
- CSN / China Securities, "Safeing human robots!
- Xinhua Zhejiang Sub-company for People ' s Insurance and Insurance (Zhejiang Branch) floating rates (three: 1 million, premium: 200-1,000 yuan per station)
- The New Wave. "The human robots will produce more than 100,000 machines a year.
- The daily economic news, "The smart robot hurt me. Who's gonna pay for it?"
- Kyoto News Network / Song Guangnet / International Finance Journal, China Insurance Industry Association: Model Provisions for the Progressive Promotion of Human Robot Insurance
- Changzhou Daily 2026-10-14 A04 Financial Edition
- Observations of the Ministry of Science and Technology of the People ' s Republic of China on accelerating the development of high-quality science and technology insurance, which strongly supports high-level science and technology self-reliance (2026-02, by name of intelligence);
- Economic Reference / Science and Technology Daily / China News Network / Workers Daily / Chinese Enterprise: Code of Practice for the Management of Human Robots throughout their Life Cycle, 29-bit code, more than 2.8 million coupons, and total number of wild trees Access
- Chinese News Weekly / Cow News: The kick of a girl by a robot
- People ' s daily newspaper / Jiang Jinnet: judgement in the case of a suicide robot in Guangdong Tween County.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.