Mifeng has turned data collection into a crowdsourcing model, embodying intelligence to welcome "data riders."

On September 23, Mifeng Technology officially launched the data crowdsourcing platform "Mifengpai", attempting to open up the collection of embodied intelligence data to the general public on a large scale.

Founded in February 2026, Mifeng Technology's business covers the collection and processing of real device data, non-ontology data, and simulation data, and it positions itself as a third-party data platform for the entire industry.

With the help of the "Mego" platform, ordinary people can apply for MEgo devices, accept data collection tasks through the App, and then the platform will settle the payment according to the valid data duration after verification.

Mifeng is also recruiting "robot trainers" and has launched a scenario data alliance with more than 50 companies and institutions in the hotel, supermarket, and manufacturing industries. The company hopes to break down data production, which was originally concentrated in data collection factories, into more real-world work and life scenarios.

Over the past year, the answer to "where does data come from" in the embodied intelligence industry has been rapidly changing. The most direct approach in the early stages was to centralize robots in data collection centers, where humans would remotely operate the robots to repeatedly perform tasks such as grasping, moving, and sorting.

However, real machines are expensive and have limited collection efficiency. It is also difficult to increase the data scale from millions of hours to tens of millions or even hundreds of millions of hours by linearly increasing the number of robots and collectors.

As a result, the industry began to separate more data generation from the robot itself.

Mifeng has chosen to further open up the production side: allowing more people to enter real-world scenarios with lightweight devices, and then having the platform process these scattered data into training data that can be used by the model.

When embodied data begins to attempt crowdsourced production, is it a rapidly scalable data network or a more complex labor business?

How is Meituan's data accounted for?

For ordinary people, participating in Mego generally involves several steps: downloading the app, renting MEgo devices, receiving tasks, completing designated operations in daily life or work scenarios, uploading the data, and earning commissions based on the duration of valid data, which can then be withdrawn.

Users can select tasks in the task hall, and ongoing tasks are subject to a time-based quota system. The MEgo device rental fee is 39 yuan per day, with a promotional discount price of 19 yuan per day. The basic reward for tasks displayed on the App is approximately 20 yuan per effective hour.

This is not the uniform hourly wage that the collectors ultimately receive.

According to the company, the compensation consists of a city-based base price and dynamic subsidies: the former is based on the income level of different regions, while the latter depends on the efficiency of task completion and data quality.

Tasks range from a few minutes to over an hour, and payment is ultimately based on the approved effective duration.

Once enough data of a certain type has been collected, the platform can stop distributing it; for scarce or difficult-to-enter scenarios, additional subsidies are used to increase attractiveness.

Currently, the open data collection tasks cover more than 20 fields, including production and maintenance, catering services, logistics and handling, elderly care and housework. Even daily operations such as counting badminton shuttlecocks in badminton halls and replenishing water cups at convenience service stations are included in the data collection scope.

"Crowdsourcing" is easily associated with variations of the food delivery rider model: platforms connect demand with scattered workers, and then dispatch them through task rules and prices.

However, the delivery difficulties of the two are different.

While food delivery services only need to deliver goods to a designated location, the data itself must meet the model's specific requirements for scenarios, actions, and quality. Furthermore, the dispersed data collectors vary greatly in their operating habits, professional skills, and working environments.

From a business model perspective, Mifeng believes that crowdsourcing can first reduce the labor costs at the data collection end.

Traditional centralized data collection requires hiring dedicated data collectors to repeatedly complete tasks in a fixed location.

The crowdsourcing model aims to embed data collection into participants' existing lives and work. The platform pays for new data contributions, rather than the cost of hiring a full-time data collector.

Correspondingly, the complexity of back-end operations will also increase.

The needs of a model company must first be broken down into tasks that ordinary people can understand, clearly defining the action steps, equipment locations, completion standards, and acceptance rules.

The platform also needs to complete basic training, equipment collection and transfer through offline service stations to minimize operational discrepancies among different data collectors.

Standardized equipment resolved some of the input discrepancies. Data that passes the initial screening still needs to undergo action segmentation, annotation, trajectory extraction, and manual sampling before it becomes usable data for the model.

Mifeng does not simply categorize data as "usable" or "unusable," but rather classifies it according to quality and usability.

Data that is clearly unusable will be removed, while the remaining data may be used for pre-training, specific task training, model evaluation, or long-tail scene learning.

According to Yao Maoqing, Chairman and CEO of Mifeng Technology, more than 95% of the data that passed the initial settlement screening can still be used after post-processing.

According to data disclosed by Mifeng, within one month of its internal testing, Mifeng Pai had 20,000 registered users and generated 13,000 task submissions. The highest-earning participant earned over 5,000 yuan in a single month.

The more dispersed the scenarios, the more complex the operations.

The timing of Mifeng's launch of its crowdsourcing platform reflects two simultaneous changes in downstream data demands: customers require increasingly larger amounts of data, and the scenarios in which these demands are being met are becoming more specific.

Yao Maoqing stated that the million-hour target is shifting from an industry supply goal to a training demand from individual clients. Some clients have already raised demands for tens of millions of hours, and some leading clients are even requiring suppliers to deliver 50,000 to 100,000 hours per week.

However, if the only goal is to produce ten times more of the same data, the value of crowdsourcing remains limited.

In the past, embodied data collection focused on home scenarios, with many teams repeatedly collecting data on tasks such as folding clothes, and this type of data has become saturated. What is becoming scarcer is new skills, new environments, and new operational processes that the model has never seen before.

In fact, Mifeng is expanding its application scenarios on both the demand and supply sides.

Mifeng Technology stated that the company holds regular production plan review meetings to summarize market orders, first determine which demands can be met by existing inventory, and then allocate the missing parts to different collection paths according to order value and resource conditions.

Yao Maoqing also mentioned that leading clients have already formed some common requirements in terms of real-world scenarios, spatial accuracy, and action semantic labels, but specific needs will continue to be refined down to different processes and actions.

Centralized data acquisition centers can replicate standard environments such as kitchens, living rooms, and desktops, but it is difficult to replicate car repair shops, hotel services, agricultural production, or specific factory production lines in a fixed location.

Finding these scenarios one by one with a dedicated team is costly and slow. Crowdsourcing, on the other hand, distributes scenario finding and data collection to people already working in different industries, making it easier to reach long-tail professional scenarios and their operational skills.

Meanwhile, Mifeng Technology hopes to leverage the "Scenario Data Alliance" to connect with various stakeholders such as factories, communities, hotels, shops, and logistics centers in advance, resolving issues related to security, authorization, and partnerships before customer needs arise.

Zhang Zhifu, Vice President of Business at Mifengpai, also mentioned that in the future, the platform may match users based on their profession and historical tasks, and allow users to submit their own scenarios. The platform will then review these submissions to determine whether they match downstream needs.

In his explanation, it's difficult for the platform's internal team alone to list thousands of worthwhile tasks in advance. Those actually working on the front lines of agriculture, maintenance, and production are much clearer about the scenarios and skills they possess.

What Mifeng ultimately wants is not a larger data collection factory, but a scenario network that can be scheduled according to orders.

This also makes operations one of the core competencies of embodied data companies.

Yao Maoqing believes that once the industry truly enters the stage of scaling up and profitability, it will "highly value operational capabilities":

The goal is to transform the data gaps of model companies into tasks, find the corresponding people and scenarios, and process highly heterogeneous data into a unified specification, so as to maintain a cost advantage while delivering on time.

More importantly, as procurement volumes increase, customers are also raising their requirements for supplier scale, quality, and delivery capabilities. Based on this, Yao Maoqing predicts that embodied data suppliers may further concentrate, and even show signs of oligopolistic trends.

According to Mifeng, supporting the production of tens of millions of hours of data requires an investment of hundreds of millions of yuan in fixed assets in data collection equipment, personnel settlement, data storage, and computing infrastructure.

The industry chain may see further specialization in hardware, operations, and data post-processing in the future, but end-to-end coverage of the entire chain is not a business that can be easily accomplished by a typical startup team.

Furthermore, for model companies that require millions or even tens of millions of hours of data, managing dozens of suppliers simultaneously means having to handle the differences in hardware, formats, and quality systems themselves.

The more suppliers there are, the higher the cost of data cleaning, transformation, and training validation will be.

In other words, the scenarios can be distributed, but the delivery interface is best centralized.

Whether it's expanding the amount of data or connecting different scenarios, the ultimate goal is not just to compete for current orders, but to secure a niche in the data production and scenario entry points before demands continue to change.

Who will pay for the next 100 million hours?

The crowdsourcing platform, Mifengpai, represents Ego's ontology-free data collection, which currently primarily serves the pre-training of embodied intelligence companies, world model companies, and general large model companies.

The model learns the relationships between objects, actions, and the environment from human interaction to expand coverage of different scenarios and skills.

Simply put, the question is whether we have seen enough models without a physical entity, and whether we can actually build a robot using real machines.

Yao Maoqing stated that if it is only to complete a specific task demo, using the target robot to collect real machine data will have a more direct effect; when the sample size is small, Ego data does not significantly improve a single task, and its value is more reflected in the zero-shot generalization ability when facing new tasks and new environments after the scale is expanded.

As models begin to make decisions in the physical world, what's needed isn't just more hours of data, but also richer scenarios, higher spatial accuracy, and more complete action semantics. The more powerful the model, the more complex the tasks it can learn and engage with, and new data gaps may emerge as a result.

At least under the current training paradigm, lack of ontology data is not a temporary substitute when there is a shortage of robots.

The variable with a longer timeframe is the data that the robot itself begins to generate.

For mature tasks such as box handling, sorting, and fixed production line operations, the robot in operation can directly record joint status, control commands, and success, failure, and abnormal situations.

These data are naturally consistent with the target ontology, and as robots are deployed on a large scale, the relative importance of human data collection in mature tasks will decrease.

For Mifeng, the more critical question is whether customers will continue to purchase from third-party platforms in the long run, even if the industry still needs a lot of data.

Once leading robotics companies establish their own installed base, scenario networks, and data feedback systems, the economics of self-collection will improve, and the data production rights for mature tasks may also shift to robotics operators.

Human data still has value, but this does not automatically equate to the continued value of third-party data collection platforms.

Yao Maoqing positions Mifeng as a "turnkey" data service.

He believes that most embodied companies still tend to concentrate their resources on hardware and model algorithm development, while outsourcing asset-heavy and operationally-intensive aspects such as equipment, manpower, scenarios, and data operations to professional platforms. Some standardized data can also be sold multiple times across customers, reducing the duplication of similar data collection.

How far human crowdsourcing can go ultimately depends on two things: how many worlds robots have yet to enter, and whether third-party platforms can find and produce this data at a lower cost than the customers themselves.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.