人形机器人在中国各大城市随处可见。它们在咖啡馆里制作浓缩咖啡,在商店里接单,拿起软饮(非常缓慢地)递给顾客。有些情况下,它们还在餐厅里准备菜肴。这些表演都大张旗鼓地进行。一个长得像人的机器人在公共场合几乎总会吸引一群围观者。人形机器人在会议和政府活动中也已无处不在。
Humanoid robots are popping up all over China’s biggest cities. They are making espressos in cafés, taking orders at shops, grabbing soft drinks and handing them (very slowly) to customers. In some cases they are preparing dishes in restaurants. These performances are conducted to great fanfare. A human-looking bot in public almost always draws a crowd of onlookers. Humanoids have also become ubiquitous at conferences and government events.
社交媒体上充斥着此类遭遇。这似乎正是目的。大批机器人并非旨在为公众服务,而是要让他们惊叹,也让世界惊叹。去年6月,中国领导人要求地方政府和国有企业到2027年初让1万台人形机器人投入实际工作。所以,更多的人形机器人即将出现。但它们仍然不会很有用。
Social media are full of such encounters. That seems to be the point. Legions of droids are not meant to serve the public, but to wow them, and the world. In June China’s leaders told local governments and state-owned enterprises to have 10,000 humanoids doing real work by the start of 2027. So more are coming. But they will still not be very useful.
中国是人形机器人生产领域的无可争议的领导者。数百家曾专注于汽车供应链的公司现在生产机器人零部件,因为机器人与汽车共用许多相似的零部件。仅用几年时间,世界上最先进的人形机器人供应链就在上海及其周边几个城市(包括人工智能中心杭州)兴起。这催生了世界上规模最大、技术最先进的人形机器人制造商,例如宇树科技,该公司去年8月成为该领域首家上市公司。
China is the undisputed leader in humanoid production. Hundreds of firms that once focused on the automotive supply chain now make parts for robots, which share many similar fittings. In just a few years, the world’s most advanced humanoid supply chain has sprung up in Shanghai and several surrounding cities, including the AI hub of Hangzhou. This has helped spawn the world’s biggest and most sophisticated humanoid-makers, such as Unitree, which in August became the first listed company of its kind.
中国公司制造了2026年全球估计生产的5万台机器人中的几乎全部。摩根士丹利银行的分析师估计,2027年的产量至少将翻一番。10万台新机器人的出现意味着,看到一台机器人,甚至喝到一台机器人制作的咖啡,的机会将大幅增加,至少对中国城市居民而言是如此。在低成本零部件供应激增的背景下,转型进入人形机器人生产的工业公司数量也在增加。到2026年第三季度,人形机器人制造商的数量可能已超过300家。更多公司将在2027年加入这场竞逐。
Chinese firms made almost all of the estimated 50,000 bots that were produced globally in 2026. The production figure will at least double in 2027, reckon analysts at Morgan Stanley, a bank. The existence of 100,000 new droids means the chances of seeing one, and even drinking a coffee made by one, will go up considerably, at least for urban Chinese. Amid the surge in supply of low-cost parts, the number of industrial firms diversifying into humanoid production is also growing. By the third quarter of 2026 the number of humanoid-makers may have surpassed 300. More will join the rush in 2027.
人形机器人的吸引力在于,它们本应多才多艺,且适合人类建造的环境。然而,在实践中,很少有人形机器人被部署到真实工作场所,因为它们完成大多数任务的能力很差。
The appeal of humanoids is that they are supposed to be versatile, and fit well into environments built for humans. In practice, however, few humanoids are being deployed in real workplaces because they are terrible at completing most tasks.
在能够发挥作用的地方,几乎总是涉及重复动作,周围环境没有任何变化。它们的笨拙在2027年不会显著改善,因为制造一台有用的人形机器人所需的数据量巨大。为这类机器人生成数据也非常昂贵。生成一小时数据的成本为500至700元人民币(75至105美元)。专家估计,要开发出能够胜任大量任务的机器人,需要1亿小时的数据。这大约需要75亿美元,并且需要数年时间才能生成。
Where they are doing useful work, it almost always involves repetitive movements, without any variation in the surrounding conditions. Their clumsiness will not diminish significantly in 2027 because the amount of data required to create a useful humanoid is vast. Generating data for this kind of robot is also very expensive. It costs 500-700 yuan ($75-105) to produce an hour of data. Experts reckon 100m hours of data will be needed to develop a bot that can excel at a useful number of tasks. That would cost roughly $7.5bn and take years to produce.
国家希望提供帮助。过去两年,一个训练中心的热潮已经开始,人形机器人在这些中心执行任务,唯一目的是生成数据。到2026年中期,约有90个训练中心已部分或全部建成,其中大多数显然由地方政府支持。未来一年将宣布更多。要了解中国的人形机器人在2027年的表现,这些训练设施将是观察的重点。但机器人要做出一杯完美的卡布奇诺——或者,确实,学会在主流社会中做任何其他有用的事情——所需的时间,将比许多人目前想象的要长得多。■
The state wants to help. A boom in training centres, where humanoids perform tasks for the sole purpose of creating data, has begun over the past two years. About 90 had been partially or fully built by mid-2026, most of them apparently backed by local governments. Many more will be announced in the coming year. For a glimpse of how China’s humanoids are doing in 2027, these training facilities will be the places to look. But it is going to take a lot longer for the robots to make a perfect cappuccino—or, indeed, to learn to do anything else that is useful in mainstream society—than many people currently imagine. ■