文｜硅碳变量，作者｜胡一刀
一边是机器人公司为了融资和上市制造订单，一边是不断降价、层层转包的数据，一位数据采集商无奈表示，“这门生意明年会是什么光景，还很难说”。
这两年，具身智能如火如荼，但不同于大语言模型，互联网有海量的数据可以采集，机器人此前基本没有现实数据，由此，数据采集和交易这一生意突然变得火热起来。可是，伴随着新一轮的上市潮，机器人的数据采集正在引发越来越多的质疑，前不久梅卡曼德创始人邵天兰的公开发声，更是直接把部分机器人企业借助数采中心包装虚假营收的乱象摆在了台面上。
一时间，数采不再是产业基建，反而成了关联交易、报表注水的“代名词”。
更令从业者不安的是，人形机器人的订单开始集体挤水分，明星企业的估值泡沫似乎也被戳破了。一旦市场动荡，数采这门生意可能更不好做了。
左手倒右手的“游戏”要结束了
在全国，不到两年时间，已经有106座人形机器人数采中心拔地而起，其中大部分数采中心是由政企共建。
也正是在这种合作下，数采中心、机器人企业及当地政府日渐摸索出一套模式：企业把机器人卖给中心，收入计入企业营收；中心招聘员工，用买来的机器人采集数据；企业再花钱把数据买回来，用于模型训练。
由此，一笔资金，在数采中心、机器人企业与当地政府中间转了一圈，三方都成功获得了成绩。
其实，这套“卖设备—回购数据”的模式本身没有太大的问题，问题在于在机器人企业奔赴上市的热潮中，数采中心成了单纯做高账面收入、推高估值的“工具”，数据有没有价值反而成了其次。而且随着做数据成了最终目的，更多的乱象自然滋生。
根据财经杂志的报道，采集商老台曾替一家机器人公司“走了一笔账”，签下约数百万元的合同。一方购买硬件，另一方购买数据，双方没有完成实物交割，只做了账务往来。他的一位同行告诉他，有一家数据公司短时间内完成多轮融资，但通过众包采集数据，把投资人的钱“合法套了出来”。
不过如今这一玩法可能即将失效。一方面，资本对具身智能企业的评估标准正在重构；一位投资人称，“以前市场热的时候，订单量、出货量、收入增速很容易直接推高估值，现在大家会追问这些收入是谁贡献的，设备有没有真正用起来，客户明年还买不买。”
另一方面，从二季度起，科创板、北交所机器人企业上会，收入真实性、关联交易合理性、毛利率可持续性，成为固定问询要点，靠修饰账面业绩冲击上市的路径基本被堵死；
一个更为关键的变化是，机器人数据采集的方式正在从依赖机器人本体的真机遥操，变为无本体采集，未来，数采员只需戴着摄像头、数据手套等设备正常工作，就能为机器人采集数据。这也意味着机器人数据采集从搭建模拟场景的数采场，将走向工厂、酒店、药店、家庭、社区等真实世界。
而如果无本体采集越来越普遍，数采中心或数采企业自然不需要花一大笔钱去购买机器人，机器人企业也就无法让它们消化自己的订单了。
具身智能“挤泡沫”，数采没有故事了？
机器人行业数据是个难题。据贵州省大数据局今年7月披露，当前我国真实场景合规数据存量约50万小时，而商业化落地需要数千万小时，缺口超过99%。
巨大的缺口，看似让数据采集存在着源源不断的需求，不愁数据卖不出去，可随着更多的数据公司、人力公司和中间商进入，数据的价格已然开始下降。老台称，去年每小时能卖200多元的数据，今年只能卖100多元。
在原来那种“卖设备—回购数据”的模式下，先投入的是数采中心或数采企业，机器人公司可以推迟回购，也可以重新压价。如今，又因为采集的方式正在从集中式转向分布式，数据采集变成了一门组织劳动力的生意，数据往更便宜的劳动力走，价格进一步被拉低了。
价格下降，交易混乱，机器人数据采集正在变成一门“苦哈哈的生意”，但更大的风险在于具身智能赛道过热，出现明显的估值泡沫，一旦泡沫破裂，机器人企业上市失败，或是拿不到融资，它们在数据采购上的需求会大大减少。
而当前，具身智能这一赛道的繁荣表象恰恰正在被外界重新审视。一面是订单掺杂水分的事实揭开了，另一面机器人进厂打工的故事也遭受质疑，很多机器人企业说是进厂，但他们并没有真的进，或者不算进。
就在前不久，福布斯直接炮轰Figure AI，称这家企业存在巨大的估值泡沫。Figure大约造出了1000台机器人，外界却不知道其中多少台真正部署、多少台正在工作，也不知道谁在持续付款。文章提及，Figure曾与宝马进行试点合作，部署了“一支执行端到端操作的机器人队伍”，但实际上，只有一台Figure机器人在工作，现场还有一台用来轮换充电。
这两年具身智能的概念炒得火热，可人形机器人到底能不能干活还是没有得到认证，而如果人形机器人始终无法干活，就代表数据没有价值，这将直接动摇机器人数据采集这一生意存在的根本逻辑。
据不完全统计，我国有近300家机器人相关公司，覆盖灵巧手、关节、电机和各类零部件，此外还有很多上市公司也在提供人形机器人解决方案。Argo Venture Partners管理合伙人熊伟铭在公开场合指出，“300家太多了。市场可能只能容纳20到30家。剩下的公司会成为大集团的一部分，可能是上市公司的一部分，或者开始转型。”
有专业人士也推测，行业洗牌预计将在2026下半年至2027年集中显现。如果到了洗牌期，机器人企业的淘汰赛可能也正是数采行业的淘汰赛。
一个更不利的消息是，在具身智能的投资浪潮里，“挖矿人”的吸引力越来越不如“卖铲人”，资本正在加速流向关节模组、力传感器、触觉感知等细分领域。当机器人企业缺钱了，数采厂辛辛苦苦采集来的数据可能更不好卖了。
别把数据采集的生意想的太美好了
这两年具身智能乘着大模型的技术浪潮和政策的引导及鼓舞，以全球瞩目的姿态站在了风口之上，其庞大的训练需求也催生了数采中心的遍地开花。我们看到，数千平方米的空间里，家居、商超、工业等场景被搬进同一座“训练场”，上万台机器人分布其间，练习物品收纳与分类、快递打包装箱、床铺整理等各种任务。
这无疑是一种重资产模式。根据中国信通院等机构近期发布的《具身智能训练场研究报告（2026年）》显示，一个数千平方米的综合性训练场，建设投入可达数千万元至上亿元，机器人采购占据较大比例，年度运营费用也可达数百万元。
前期投入大，如果后续收购价格偏低，或者数据不符合要求，设备、人力和调试投入就可能难以收回。设备即使能够转售，也可能承受较大折价。
如今随着具身智能与数据采集之间“左手倒右手”的套路被披露，一个同样的问题也指向数采，即这些重金采集出来的数据，市场上到底有没有外部买家？尴尬的事实是，绝大多数数采厂的模拟场景都集中在叠衣服、搬箱子、包裹分拣等少数通用任务中，数据重复度极高，即使是机器人企业，也不想要了。
市场正在向不需要依赖本体的分布式采集要答案。
毫无疑问，当所有人都可能同时成为戴着摄像头、数据手套的数采员，一边工作，一边为机器人采集数据，由此带来的数据是海量的，这有可能填上具身智能巨大的数据缺口。更关键的是，数据采集从模拟场景到真实场景，采集出来的数据也为头部大模型厂商及世界模型创业公司所需，这将为其带来更多真实的买家。
近期，英伟达首席研究科学家、具身自主研究团队负责人范麟熙（Jim Fan）在红杉AI Ascent 2026大会上提出，用WAM（世界动作模型）取代VLA模型，用UMI和Ego数据取代真机数据。巨头的带动，让无本体采集越发受到关注。
只是随着数据采集正在变成一门组织劳动力的生意，层层分包下，一个更加混乱的数采“江湖”也在形成。
一位数据代理商透露，一个头部数据采集公司，融资金额几十亿元，然后在山东、河南一些地方开始搞「进村采数据」，让一个村的老太太、大姨戴着 PICO 的VR设备做切菜、做饭等任务，一小时一单给5块钱。离谱的是，有一线人员反馈，组织数据采集的中间商最后直接跑路，连这点劳务费都没结。
价格混乱、数据倒卖、采集设备不成熟…这一系列的问题在分布式采集模式的发展初期是不可避免的，可更根本的问题还是这些海量的数据到底能不能用、有多少能用来训练。
而抛开商业考量，这种数据采集存在一个潜在的风险：人类向AI售卖真实的劳动数据，让机器人变得更智能，这些机器人以后很可能再来取代人类。如果这种不安和焦虑逐渐放大，谁还愿意为机器人提供“养料”呢？
拼产能、拼订单、拼落地，具身智能的赛道上，整机厂商仍在狂奔，但理性的声音越来越大。当狂热消失，这一产业上大大小小的企业都应该考虑下一步该往哪走。
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。
It's a vinyl silicon carbon variable.
On the one hand, robotic companies make orders for financing and listing, and on the other hand, data from declining prices and layers of subcontracting, one data collector said, “It is hard to say what this business will look like next year”.
In the past two years, the business of data collection and trading has suddenly become hot, unlike large-language models, where big amounts of data can be collected on the Internet and where real-life data have been largely unavailable to robots. However, along with a new wave of listings, robotic data collection is giving rise to more and more questions, as the public voice of the founder of Mercamand, Shang Tenland, was recently put on the table by some robotic firms, with the help of digital centres, to wrap up false collections.
At one point, it was no longer the building blocks of the industry, but rather the “noun” for the association transaction, the statement of water.
Even more disturbing to practitioners is the fact that orders for human robots have begun to squeeze water collectively, and the valuation bubbles of star companies appear to have broken. Once the market is volatile, this business may be even worse.
The game is over with the left hand to the right.
Nationwide, in less than two years, 106 centres for the collection of man-made machines have been uprooted, most of which have been built by political enterprises.
It is also with this cooperation that the centres, robotics and local governments increasingly find a model: businesses sell robots to the centres and account for their revenues; centres recruit employees to collect data from the machines they buy; and businesses buy back the data for modeling.
As a result, a single sum of money was turned around between centres, robotics and local governments, all of which were successful.
In fact, the “selling of equipment-repurchase data” model itself is not a big problem. The problem is that, in the boom of robotics, digital centres become “tools” for simply high-account income and high-value valuations, and data are second. And as data become the end, more chaos breeds.
According to the financial magazine, the collector's office had “disappeared a account” for a robotic company and had contracted about millions of dollars. One party purchased hardware, the other purchased data, and the other party did not complete the physical delivery and only made a transaction. One of his colleagues told him that a data company had completed multiple rounds of financing in a short period of time, but had collected data through crowdsourcing, “legitimately set up” the investor's money.
But now this game is about to expire. On the one hand, the capital assessment criteria for smart enterprises are being reformulated; one investor states, “It was easy to push up the valuation directly when the market was hot, the volume of orders, the volume of deliveries and the speed of income, and now you will be asking who contributed these revenues, whether the equipment actually worked and the customers will buy it next year”.
On the other hand, starting in the second quarter, B.O.C. and B.O.O.C.C., with income authenticity, the legitimacy of related transactions, and the sustainability of the M.A.R. rate, have become fixed points of inquiry, and the path to the market through book-keeping impact has largely been blocked;
A more critical change is that robotic data are being collected in a way that is moving away from real robotic telemetry, which relies on the robotic body itself, and that, in the future, a few pickers will be able to collect data for robots if they are properly equipped to wear cameras, data gloves, etc. This also means that robotic data collection will go from the mining fields where simulation scenarios are built to the real world of factories, hotels, pharmacies, families, communities, etc.
And if unincorporated collection is becoming more common, and it is natural that digital centres or firms don't have to spend a lot of money buying robots, robotic firms will not be able to absorb their own orders.
Smart Squeezing Foam. No more stories?
The robot industry data is a problem. According to the Grand Data Bureau of Guizhou last July, the current stock of real scene compliance data in the country is approximately 500,000 hours, while commercialization takes tens of millions of hours, representing a shortfall of more than 99 per cent.
Large gaps, which seem to create a constant demand for data collection, cannot be sold, but prices have begun to decline as more data companies, manpower companies and brokers enter. According to the station, more than 200 data per hour were sold last year, and only 100 this year.
Under the original “sale of equipment-repurchase data” model, the first input was from a digital centre or company, and robotics companies could postpone repurchase or re-press prices. Today, again, because of the shift from a centralized to a distributed approach to collection, data collection has become a business of organizing labour, moving to cheaper labour, and prices have been further lowered.
Prices are falling, transactions are confusing, robotic data collection is becoming a “hard-ha-ha-ha business”, but the greater risk lies in overheating the smart track, the emergence of visible valuation bubbles, and their need for data procurement will be reduced significantly if the bubble breaks, robotics fail to market or cannot be financed.
At the moment, the appearance of prosperity on the intelligent track is being re-examined. On the one hand, the fact that the order was mixed with water, on the other hand, the story of the robot working in the factory was questioned, and many robotic companies said they were going in, but they didn't really, or they didn't.
Just recently, Forbes directly shelled Figure AI, claiming that the company had a huge valuation bubble. Figure has built about 1,000 robots, but the outside world does not know how many of them are actually deployed, how many are working and who is paying for them on a continuous basis. The article mentioned that Figure had worked with the BMW on a pilot basis to deploy “a team of robots to perform end-to-end operations”, but that in practice only one Figure robot was working and that there was another on-site to rotate chargers.
The fact that these two years of smart thinking have been hottening up, that human-shaped robots are working or are not certified, and that if human-shaped robots remain unable to work, they represent data that are of no value, will directly undermine the very logic of robotic data collection.
According to incomplete statistics, there are nearly 300 robot-related companies covering smart hands, joints, electric appliances and various parts, in addition to many listed companies that are providing human robot solutions. Argo Venture Partners' Management Partner, Kui Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Wei Markets may accommodate only 20 to 30 households. The remaining companies will become part of a large group, possibly part of a listed company, or begin to transform.”
Professionals also speculate that the trade shuffle is expected to be concentrated in the second half of 2026 to 2027. If the shuffle is over, the robotics phase-out competition may be the one in the digital industry.
On a more negative note, “minediggers” are becoming less attractive than “sellers”, in a wave of smart investments, and capital is accelerating to subdivisions such as joint modulations, powerful sensors, tactiles, etc. When robotic enterprises are short of money, the data collected by the mills may be harder to sell.
Don't think the business of collecting data is too good.
The two years of smartness, guided and inspired by the technological wave and policy of LLM, have been at the forefront of the global spotlight, and its enormous training needs have led to the flourishing of centres. We have seen thousands of square metres of space in which home, business, industry, etc. are moved into the same “training grounds”, where thousands of robots are distributed, practicing a variety of tasks, such as receiving and sorting items, packing boxes for express delivery, bedmaking, etc.
This is undoubtedly a heavy asset model. According to the recent publish study on Smart Training Sites (2026) of China’s Intelligent Training Campus, a thousands of square metres of integrated training grounds can be built with tens of millions to hundreds of millions of yuan in investment, with a large share of robot purchases, and annual operating costs of millions.
The high level of pre-commitment may make it difficult to recover equipment, manpower and commissioning inputs if the prices of subsequent acquisitions are low or data do not match requirements. Even if the equipment can be resold, it may bear a higher discount.
Now, with the disclosure of the “left-hand-to-right” route between intelligence and data collection, a similar question, namely, whether there are external buyers on the market, is the data collected from these heavy gold? The embarrassing fact is that most of the mining plants ' simulations are concentrated on a few generic tasks, such as folding clothes, moving boxes, sorting packages, etc., with very high levels of data repetition, even for robotics.
Markets are asking for answers to distributional collections that do not need to rely on the body.
There is no doubt that when all people can simultaneously be digital collectors with cameras and data gloves, while working and collecting data for robots, the resulting data is big, and this may fill a huge data gap with intelligence. More critically, data capture, from simulation to reality, is also needed for head LLM manufacturers and world model start-ups, which will bring more real buyers.
Recently, Jim Fan, leading research scientist and head of an autonomous research team in Ingweida, proposed to replace the VLA model with WAM (World Action Model) and real data with UMI and Ego data at the Auscent 2026 Conference. The power of giants has led to increased interest in inorganic collection.
Just as data collection is becoming a business of organized labour, a much more chaotic number of “gang and lake” is forming under layers of subcontracting.
A data agent revealed that a head data collector company financed billions of dollars, and then started to collect data in the villages in Shandong and Henan, where the village's old wife and sister-in-law wore PICO's VR equipment to cut vegetables, cook meals, etc. for $5 an hour. Paradoxically, there is a line of feedback, and the middlemen who organize the data collection end up running directly, even at no cost.
Price disorder, data re-sold, collection equipment premature... This set of questions is inevitable in the early stages of the development of distributive collection patterns, but the more fundamental question is whether or not these big amounts of data can be used and how much can be used for training.
Beyond commercial considerations, there is a potential risk that this kind of data collection will sell real labour data to AI to make robots more intelligent, which are likely to replace humans later. If this anxiety and anxiety are gradually magnified, who else would be willing to provide the robot with “food”?
The whole machine continues to run, but the voice of reason is growing. When fanaticism disappears, companies big and small in the industry should consider where next step should go.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.