中国正在训练数以万计的人形机器人：《经济学人》双语精读
本文为《经济学人》2026 年 8 月 29 日 business 栏文章的中英双语精读。中文完整覆盖原文论证与数据，英文保留各段关键原句并附注释，末尾附词汇与表达笔记。
导读
中国人形机器人正在把竞争者甩在身后——而且是字面意义上的「甩在身后」。文章的核心判断是：中国几乎已经锁定了机器人数据采集的全球领先地位，但这份领先是用巨额补贴买来的，而商业化在本十年内都不会到来。
一条贯穿全文的技术线索：硬件已经解决，瓶颈在数据和软件。而机器人的数据，昂贵、脏、效率低，和大语言模型吃互联网文本的处境完全不同。
一、跑得比博尔特快，然后撞墙
On August 17th one of their number, called Superman, clocked up a running speed of 12.66 metres per second, beating the previous record set by Usain Bolt, before crashing into a wall.
8 月 17 日，一台名为「超人」的机器人跑出 12.66 米/秒，打破了尤塞恩·博尔特保持的原纪录——然后一头撞在墙上。这场「胜利」为它的制造商宇树科技（Unitree）添了一把火：两天后该公司在上海上市，股价暴涨 460%。
Yet as the ignominious end of Superman's running feat demonstrates, the technology still has a long way to go.
然而，「超人」这场奔跑的可悲结局恰恰说明：这项技术还有很长的路要走。
本段数据：今年中国厂商预计销售 5 万台人形机器人，是去年的三倍多。硬件基本已被攻克，软件则是另一回事。
二、模型的差距：几十亿 vs 数千亿参数
The models powering the current generation of humanoids possess a few billion "parameters"; to replicate the functioning of the human body, they will need hundreds of billions.
驱动当代人形机器人的模型只有几十亿「参数」；要复现人体的运作方式，需要数千亿。
Gathering the data necessary to teach a robot to identify a glass, pick it up and fill it with coffee is much harder.
而要教会机器人识别一个杯子、拿起来、倒满咖啡，所需数据的采集要难得多。
The bot must understand everything from its spatial position to the brittleness of the glass and the viscosity of its contents.
机器人必须理解一切：自己所处的空间位置、玻璃杯有多易碎、杯中液体的黏度是多少。
要点：参数规模上，人形机器人的基础模型最终会与大语言模型相当。但大模型可以用互联网上堆积如山的数字文本训练，机器人不行——它要靠真实世界的经验，或对真实世界的精密复刻。
三、真机数据 vs 仿真数据
Humanoids are currently fed a hotchpotch of data, often from low-quality simulations. Data collected from the real world are far more useful—and now highly sought after.
目前人形机器人被喂的是大杂烩般拼凑的数据，且常常来自低质量的仿真。而来自真实世界的数据要有用得多——如今也成了各方争抢的对象。
The most valuable form is "real machine" data, which are generated by the actions of a robot, either when it is controlled remotely or when it is carrying out tasks by itself.
最有价值的一类是「真机数据」：由机器人自己的动作产生——无论是被远程操控，还是自主执行任务。
Another important source comes from recording humans as they complete tasks. Known as "egocentric" data, this information is gathered by equipping a person with a headset, haptic-sensor gloves and other kit that records their movements.
另一个重要来源是录制人类完成任务的过程。这类信息被称为「第一人称数据」（egocentric data），采集方式是给人戴上头显、触觉传感手套等记录动作的装备。
四、53 座训练中心
At the sprawling facility in the central Chinese city of Wuhan, up to 100 robots spend hours each day observing and mimicking humans.
在华中城市武汉的一座庞大设施里，最多有 100 台机器人每天花上数小时观察并模仿人类。
At one of the dozens of work stations, a young man behind a counter is moving his hand, gloved in digital sensors, in a circular motion as if he were pouring hot water over coffee. Next to him stands a robot that is mimicking his motion but spouting real water from a kettle over a funnel of coffee.
在众多工位中的一个柜台后面，一个年轻人戴着装了数字传感器的手套，画圈般地晃动手，像是在往咖啡里注热水。旁边站着一台机器人，模仿他的动作，用水壶把真正的水浇进漏斗装的咖啡里。它的手腕和手指移动的方式，诡异地像人类。
The centre in Wuhan is one of 53 facilities in China built especially for gleaning the secrets of human movement.
武汉这家中心，是全国 53 座专门为「发掘人体运动奥秘」而建的设施之一（据咨询公司 Interact Analysis 数据）。其中多数建于过去两年内，另有 34 座在建或已规划。上海在建的一座将容纳 1000 台人形机器人；浙江沿海一座小城已有六座训练中心。
(Until now foreign journalists have not been allowed to visit these centres.)
（迄今外国记者尚未获准参观这些中心。）
五、昂贵的低效
Training centres require a large amount of space and many people to run them, as well as a legion of cutting-edge humanoids, which can run to 100,000 yuan ($15,000) apiece.
训练中心需要大量场地、大量运营人员，以及一支庞大的前沿人形机器人队伍——每台售价可达 10 万元（1.5 万美元）。
And for every eight hours of coffee-pouring, for example, the droid at the café station produces just three hours of useful data.
以咖啡工位为例，每倒八小时咖啡，这台机器人只产出三小时有用数据。
The rest is spoiled, mostly by errors that engineers do not want the bots to learn. This form of training thus costs 500-700 yuan per hour, reckons Corey Chan of HSBC.
其余的都废了——主要是工程师不希望机器人学会的那些错误。据汇丰银行 Corey Chan 估算，这类训练的成本是每小时 500 至 700 元。而据一项估计，要开发出一台能执行多种日常任务的机器人，需要 1 亿训练小时。
六、政府买单
Around four-fifths are state-backed. Many are collaborations between private enterprises and local governments, which buy humanoids and sell the data produced back to the companies.
约五分之四有国家支持。许多是民企与地方政府的合作：政府买入人形机器人，再把产出的数据卖回给企业。
Estimates vary but in the first half of 2026 up to 70% of the humanoids made in China may have gone to these training centres, adding up to around 13,000 machines.
各方估算不一，但 2026 年上半年，中国产出的机器人中可能有多达 70% 流入了这些训练中心，合计约 1.3 万台。
By building one, local officials can shore up demand for robots that are being manufactured in their area.
建一个训练中心，地方官员就能为本地制造的机器人托住需求。
要点：这是一种产业政策式的循环补贴——地方要 GDP 与产值，数据中心消化了本地产的机器人，机器人厂商则拿到近乎免费的数据。对宇树这样的企业而言，账本因此好看了很多。
七、穿着设备做兼职
Chinese robot-makers are finding ways of collecting egocentric data from the country's large manufacturing workforce—as well as its many unemployed youngsters.
中国机器人厂商正在想办法，从该国庞大的制造业劳动力——以及大量失业的年轻人——身上采集第一人称数据。
A cottage industry has emerged of vendors making wearable gear for that purpose. Some hope young people will wear the kit while at home as a part-time job.
一门制作此类可穿戴设备的小产业已经出现。有人希望年轻人在家把这套装备当兼职来穿。
JD.com, an e-commerce firm, said earlier this year that it has launched a project to pay 100,000 of its staff and 500,000 others to track their own movements.
电商公司京东今年早些时候宣布，启动一个项目，付费 10 万名自有员工和另外 50 万人来追踪自己的动作。在江苏一座据报道获得国家资助的设施里，该公司付钱让人们在仿真杂货店等环境中执行任务，未来两年计划收集 1000 万小时真实人类场景数据。
Egocentric data generated by humans are far less useful than real machine data, owing to the differences in how human limbs and robotic ones move in practice.
人类产生的第一人称数据远不如真机数据有用，因为人的四肢和机器人的四肢在实际运动方式上存在差异。
Still, China gained an early edge in image-recognition technology by employing large groups of people to tag pictures.
不过，中国在图像识别技术上曾靠雇佣大批人力给图片打标而取得先机。一位机器人公司工程师认为，同样招数可以再来一次。
八、美国的另一条路
Since its initial pop, Unitree's share price has slumped by 30%.
自上市首日暴涨以来，宇树的股价已回落 30%。
Tesla, America's biggest producer of humanoids, is said to be training robots in its own warehouses, which saves it the cost of constructing standalone training facilities.
美国最大的人形机器人生产商特斯拉，据说在自有仓库里训练机器人，从而省去了单独建造训练设施的成本。
And since these days America has few factory workers who can be handed gear for egocentric collection, a number of firms are reportedly paying workers in poor countries where wages are lower than in China to don headsets and haptic gloves as they work.
而如今美国工厂工人所剩无几，无从人人发一套设备去采第一人称数据，于是据报有多家公司改为付费给工资比中国更低的贫穷国家工人，让他们干活时戴上头显和触觉手套。
Over the next few years it is set to produce and sell hundreds of thousands of costly but mostly useless machines in the hope that it can make cheap and useful ones in future.
未来几年，中国人形机器人产业将生产并卖出数十万台昂贵却基本无用的机器，寄望于将来能造出便宜又好用的。
九、结论：领先，但要付多少代价
Given the current state of the technology, and the time that will be required to improve it, a commercial market for humanoids is unlikely to emerge this decade.
就当前技术水平、以及改进所需的时间而言，人形机器人的商业市场在本十年内不太可能出现。
Many will remain dim-witted—and continue charging into walls.
其中很多将始终愚钝——并继续一头撞进墙里。
最终数字：摩根士丹利估计，到 2030 年中国每年可能卖出近 45 万台机器人。
词汇与表达笔记
| 表达 | 释义 |
|---|---|
| clocked up a speed | （口语）跑出/达到某个速度。新闻英语常见搭配 clock up 12 metres per second |
| ignominious | 不光彩的、可悲的。此处形容「超人」的结局，带讽刺意味 |
| crash into a wall | 撞上墙。机器人的失败常与「炫技」形成反差 |
| a hotchpotch of data | 大杂烩般拼凑的数据。hotchpotch 源自「杂烩锅」，英语里专指勉强凑起来的东西 |
| real machine data | 真机数据。由机器人自身动作（遥操作或自主执行）产生的数据 |
| egocentric data | 第一人称/自我中心数据。穿戴设备录下人自己的动作作为代理数据 |
| haptic-sensor gloves | 触觉传感器手套。能捕捉手部动作与力度的设备 |
| apiece | 每件/每个。正式书面语，财经报道常见 |
| a legion of | 一支庞大的（队伍）。legion 原为古罗马军团 |
| reckons | （据某人的）估计。读作 /ˈrekənz/，第三人称单数动词 |
| coast / pop | 股价首日暴涨。pop 是俚语，比 surge 更口语 |
| charge into walls | 撞进墙里。文章结尾的收束，呼应开头「超人」撞墙的细节 |
| slump by 30% | 回落 30% |
| warehouse training | 用自有仓库当训练场。省下单独建设施的成本 |
| cottage industry | 家庭式小作坊产业；小本经营的行当 |
可直接借用的表达
They have largely mastered the hardware behind the machines. The software, however, is another matter.—— 硬件已解决，软件另说。这是科技报道里非常典型的转折句式。X is set to produce ... in the hope that it can make ...—— 未来将生产 X，指望能做出 Y。写产业分析时很好用。Given the current state of the technology, ... is unlikely to emerge this decade.—— 就当前技术水平而言，…… 本十年内难以出现。克制判断的规范说法。
原文：China Is Training Up Thousands Of Humanoid Robots, The Economist, 2026-08-29, business 栏。英文引句为各段关键原句，中文为对照翻译与要点整理。