Rural aunties act as "mentors" for robots

Perhaps they themselves have no idea that they have already become the trendsetters of the era.

Peeling peanuts, packing them into bags, then weighing them on a scale. In rural areas of Henan Province, elderly women are doing farm work they are extremely familiar with. If there is any difference between the present and the past, the only one is that everyone wears a wearable data collection device on their head. This certainly brings a little inconvenience, but they are willing to do it.

These elderly women have lived in rural areas all their lives, and they have no other sources of income except the harvest from their fields. But now, they can get a daily salary of 80 yuan for working 8 hours a day, which adds up to more than 2000 yuan a month.

Just like college students who work in high-end office buildings and industrial parks, these elderly women are now practitioners of embodied intelligence, and their official job title is data collector.

Earlier this year, NVIDIA released a major breakthrough called "EgoScale". When researchers fed more than 20,000 hours of human first-person perspective videos to the large model, they found that as long as the number of human demonstration data keeps increasing, the operation capability of robots can be steadily improved.

In the past, data collection in the embodied intelligence industry mainly relied on ontology-based collection, which required obtaining data through methods such as manipulating real machines. After the release of this achievement, non-ontology collection has also become a mainstream data collection method, which can be completed by real people wearing data collection terminals.

Compared with ontology-based collection, non-ontology collection greatly lowers the threshold for data collection. It does not require the cooperation of expensive robot ontologies, and collectors do not need professional educational background. It barely has any threshold, and has become an occupation open to the public.

Virtually overnight, this industry that was originally defined as high-tech has rapidly attracted the influx of grassroots groups.

01

Zhou Yan is engaged in human resources outsourcing, delivering labor to various factories. In the past, winter was his least favorite season, because at this time of the year, factories generally no longer recruit additional workers, and the performance of his company would decline for several consecutive months just like the temperature, and he could only wait for the resumption of work in the next year to make gradual plans.

But Zhou Yan is not anxious this year. At the beginning of this year, he started a side business of data collection, and now his work focus has even shifted to this side business.

Similar to those elderly women, in Zhou Yan's perception, what he is doing now is actually not much different from what he did in the past. The essence of data collection is also "human power".

Now he is in contact with a commercial collection project. The Party A has a collection demand of 500,000 hours, which is absolutely a large order. However, because the two parties have not reached an agreement on the price, the contract has not been signed yet.

But Zhou Yan is not in a hurry at all. According to the human resources he has, it is no problem to take on the demand of 200,000 to 300,000 hours. Very few teams on the market can achieve this, so he thinks there is still room for further negotiation on the price.

Even when the contract has not been finalized, he has already started preliminary communication with other suppliers for subcontracting cooperation.

Xu Yang used to be engaged in field promotion. With different experience, he has a different playing style from Zhou Yan, and he is good at expanding commercial resources. Xu Yang has enough experience in communicating with small merchants, so it is not difficult to get them to cooperate in completing data collection work.

There are particular considerations in selecting regions. Xu Yang will first exclude provincial capitals, because merchants who open stores there come from all over the country, and communication is not easy. Unlike small counties where almost all shop owners are locals, it is easier to reach cooperation.

For quite a number of merchants, they can get an extra income just by wearing a device, so they are easy to be persuaded. But there are also some merchants who are unwilling to participate in data collection due to reasons such as insufficient trust or reluctance to go through troubles.

For the latter, Xu Yang also has his own solutions. For example, he will pay wages to hire employees to work in these stores, so that merchants can get a "free" labor force, and Xu Yang can also collect the data he wants.

However, the workload of doing data collection through field promotion is actually very large. According to the requirements of the project Xu Yang received, only about a hundred hours of data can be submitted for one scenario, and the business license credit code of the merchant needs to be submitted for verification.

This means that the collection time for each store is only about one week, and a collection project of tens of thousands of hours even needs to expand hundreds of scenarios. Xu Yang's team is still on the way to expand merchants every day.

Zhang Qingyuan took the third path. He chose to become a city partner of "Mifengpai". This is a crowdsourcing platform that releases collection tasks for more than 20 scenarios including manufacturing, retail, catering and other fields. Registered users can claim and complete the tasks as needed, but the premise is to rent the collection equipment designated by Mifengpai.

Zhang Qingyuan's job is to conduct precise drainage through social platforms, and distribute these collection devices all over the country.

Zhang Qingyuan used to work in the loan industry. But with the official implementation of the new regulations on loan assistance in September, loan intermediaries experienced a major reshuffle. Coincidentally, Mifengpai was officially launched in the same month, and Zhang Qingyuan successfully switched careers.

02

After receiving the project from the source company, Zhou Yan will subcontract or distribute the project to others.

Every morning, he will start a live broadcast on Douyin, and look for potential "suppliers" through mic linking. After establishing contact in the live broadcast room, Zhou Yan will ask potential suppliers to fill in the warehousing application form, with very high efficiency.

This practice is very common in the data collection industry.

Wang Zhen's family is engaged in the real industry, and he himself entered the technology industry and established a data collection company named Cervo Technologies.

At the very beginning, Wang Zhen just wanted to use the existing factory scenarios of his family to do data collection. Now the order volume of his company is doubling every month. Although he is also recruiting employees urgently, if he still does all the collection by himself as before, he cannot digest these orders at all.

So a few days ago, Wang Zhen decided to transform into a platform company and outsource the orders in his hands. He hopes that his company will become a stable supplier for leading model companies in the future.

Wang Zhen has overseas study experience. Through the resources of his classmates, he can easily get access to some first-hand procurement projects abroad. The prices offered by these projects are generally above 100 yuan, and the price for labeled data will double.

However, there is a big price difference after these projects flow into China. Xu Yang got an external list of embodied data collection projects in September. The unit price of some projects has dropped to 20 yuan, and the highest unit price of the projects is only 80 yuan. Xu Yang calculated by himself that the unit price must reach at least 30 yuan to avoid losses.

After the project is subcontracted layer by layer, each layer will draw its own profit, and it is normal for projects in this industry to be subcontracted two or three times.

This also means that the profit margin left for the actual execution team is very small. But Wang Zhen thinks it is not surprising, which is a typical feature of the rapid volume growth stage.

The cost of a project actually consists of two parts: one is the equipment cost, and the other is the labor cost. The former is a rigid cost with little room for adjustment. Therefore, to squeeze out more profit margin, the project needs to be pushed to regions with lower labor costs.

After Xu Yang entered the industry, his biggest feeling is that it is not easy to make money or lose money in this industry.

Around Xu Yang, there is currently only one friend who is still losing money. His friend is among the first batch of people to enter this industry. At that time, the collection scenarios were mainly built from scratch. Xu Yang's friend formed a team in Beijing and rented a house to build scenarios, but he has not even earned back the rent so far.

A consensus has long been formed in this industry: if you want to make money, you need to stay away from cities, especially high-tier cities. This consensus has also forced the data collection business to sink into regions such as counties. Grassroots groups including small town youth, full-time housewives and rural elderly have become the core force of data collection.

Of course, another key point why this job can cover grassroots groups is that although it has no requirements for education and technology, it has requirements for physical strength, and people need to have the tenacity to stick to boring work.

A month ago, Xu Haiyan became a data collector, and she is mainly responsible for collecting data of home scenarios. When working, Xu Haiyan needs to wear a gripper on her hand and a device on her head. Each device weighs about 0.6kg, and she has to wear them for seven or eight hours, which makes her very tired after a whole day's work.

While the body is tired, it is also a test for the mind. During the collection period, she must ensure that her hands do not appear in the frame, and she cannot talk to other people, let alone play with mobile phones. Each action needs to be recorded 50 times, which means that Xu Haiyan may need to repeat the action of making the bed for two or three consecutive days.

What makes Xu Haiyan even more collapsed is that after working all day, the effective duration is only four hours, and the effective duration rate is only about 50%, while her salary is linked to the effective duration.

Xu Haiyan has asked herself more than once whether it is necessary to persist. After experiencing this job in person, she finally understands why the turnover rate of this position is so high.

Zhao Feifei's full-time job is to do women's federation work in rural Shanxi. A month ago, she began to organize women in the village to do data collection, wanting to find a way for everyone to make money without going out to work, so as to subsidize their families. But a few days after the devices were distributed, they were suddenly taken back.

After communicating with peers, Zhao Feifei found that this situation is not uncommon. Some projects have very long subcontracting chains, and changes in Party A's demand rules cannot be synchronized to the execution team in time. Of course, up to now, she still has not figured out why the cooperation was interrupted.

All in all, it is not easy to fall in love with this industry.

03

A large amount of data demand will continue for another two or three years, which is Wang Zhen's judgment on this industry.

In fact, this is also the most concerned issue for practitioners, because they clearly feel that it is no longer easy to make money in this industry.

At present, the data collection teams on the market are taking two paths: one is to wait for orders and complete collection tasks according to requirements, and the other is to sell data, packaging and selling data to ontology or model companies by duration. However, the second path has been very difficult to take now.

When Xu Yang first entered the industry, the data sold in this packaged way could still be sold for 33 yuan per hour. But now it has basically dropped to less than 5 yuan, and it is generally around 3 yuan.

As for how far the first path can go, no one can tell clearly so far. Although the embodied industry has a data bottleneck, it does not mean that the demand of the data industry will continue to increase.

Wang Zhen feels that there are not many companies on the market that have both funds and graphics cards to do pre-training, and almost all data collection orders come from overseas. Xu Haiyan also confirms this statement, and the company she works for receives almost all overseas orders.

China has a complete industrial chain, a large number of scenarios available for embodied intelligence training, and a large amount of achievable data, which is our advantage. However, the cross-border flow of data is strictly regulated by laws and regulations. According to the experience of China's first cross-border transaction of embodied intelligence data, relevant data needs to complete hierarchical assessment and filing before leaving the country to ensure that the whole chain is compliant, safe and controllable.

But these issues are still too advanced for grassroots data collectors. In contrast, what they are more anxious about is where the next order is.

Zhou Zhou recently received a dialect collection project, which needs to collect dialects from five regions: Chongqing, Jiangsu, Jiangxi, Shandong and Gansu. Compared with the projects Zhou Zhou received before, this project is already very high-quality.

The collected content includes two types: single-person and two-person. For the two-person type, two people use dialect to have free conversations based on specific topics. They can get 280 yuan after talking about 5 topics for 30 minutes each. In addition, this project has very low requirements for equipment. As long as there is a mobile phone, it can be completed, and there is no need to pay extra cost to rent collection equipment, which belongs to a project with relatively high profit.

Moreover, as long as there is no obvious situation such as reading according to the manuscript during the process, it will not be judged as invalid. According to past experience, the effective duration ratio of Zhou Zhou's team in doing this kind of project can reach more than 98%, which is a project with a high audit pass rate rarely seen in the industry.

But Zhou Zhou is still very anxious. Calculated according to the duration agreed in the contract, this project can only be done for a maximum of one and a half months. After it is completed, it is hard to say what kind of orders can be received, or even whether orders can be received at all.

Of course, there are some practitioners in the industry who make quick money, but more practitioners hope that data collection can have a more long-term future.

Wang Zhen sees that there are two mountains waiting to be crossed in front of the data collection industry now: one is volume, and the other is quality. Now thousands of practitioners join the data collection industry every day, but it is still far from the volume that can get the training results expected by the market. In the early stage of the industry, all kinds of data are needed, but what is scarce now is data of real scenarios such as medical treatment.

He judges that the data required by data collection projects on the market in the future will become more and more precise and more and more vertical. And when a certain volume is reached, it will also be possible to determine which kind of data is effective. In this process, new opportunities will naturally emerge continuously.

Xu Haiyan also maintains a positive attitude towards this industry, because she has felt tremendous changes just one month after she joined the job.

The new project received by the company she works for can be completed with only one head-mounted data collection device. The hand collection can be done with natural movements, no equipment needs to be worn, which significantly reduces the burden. And with the upgrade of equipment, Xu Haiyan no longer needs to deliberately slow down her movements to prevent the picture from being blurred due to too large or too fast movement range.

More importantly, the effective duration ratio collected by the new device has also been greatly improved. To complete four hours of effective duration, plus preparation work, Xu Haiyan only needs to work five hours a day.

While she is busy with the tasks, many ideas are active in Xu Haiyan's mind, such as what kind of changes the life clips of ordinary people like them will bring to the embodied intelligence industry.

This article is from WeChat Official Account "FunTalk" (ID: iFuntalker), written by the FunTalk team, and authorized for release by 36Kr.