Nearly a hundred players flood into embodied data: with total funding reaching 4.47 billion yuan in a year, who can truly profit from "data selling"?

Lin Fangzhou, from Non-Lab

In Chenzhou, Hunan, a China Mobile service outlet has been branded as an "Embodied Data Collection 5S Store". Ordinary customers can get a set of grippers, gloves, and a head-mounted camera. After a simple training session, they can collect robot training data while doing housework.

The first batch of 1,000 deployed devices can collect 1 million hours of data annually at full capacity. I can almost hear the store's clever little plan: collect data while grabbing public attention—even 4A ad agencies should learn from this. (doge)

There are many other similar "tricks" for embodied data collection: some offer free home cleaning services to gather data (you're welcome to come to my place), some turn data collection into VR games, and others connect robots to the internet so collectors don't need to go to data factories—they can remotely "cloud-operate" the robots.

Embodied Data Collection

However, just laugh off the examples above—collecting qualified data is far from easy. The reason these "tricks" keep popping up is simple: robots are starving for data.

Everyone is now going all out to collect data, but few have fully mapped out the entire industry landscape.

Qbit conducted an incomplete survey of 97 domestic embodied data players, among which 70 focus on data collection, and 27 work on data infrastructure.

In the past year (July 1, 2025 to July 1, 2026), 15 independent embodied data service providers—companies that don't build robot bodies, don't develop models, and only focus on data—have raised a total of about 4.47 billion yuan in financing.

Given the current frenzy of capital for embodied intelligence, this number is actually not that high. Qbit previously calculated that in the first half of this year, companies focused on the "brain" side of embodied intelligence raised 22.3 billion yuan in just six months.

To help you understand the embodied data industry, we've summarized the following 10 key industry realities.

How is data collected?

Reality 1: There are four main technical routes for data collection, and the cross-route collection track is the most crowded

Currently, mainstream embodied data collection can be divided into four technical categories:

Teleoperation of real robots: Humans control physical robots to perform tasks, simultaneously collecting motion, state, and sensor data.

Body-free collection: Humans directly demonstrate actions, which are captured through motion tracking, gripper mapping, first-person cameras, and other devices—no robot required.

Simulation synthesis: Mass-generate robot interaction data in virtual environments for model training.

Internet video distillation: Extract human motion knowledge from online videos and convert it into data that embodied models can learn from.

Among the 70 data collection companies/platforms in Qbit's incomplete survey, 30 adopt multiple collection routes at the same time, accounting for 43%—for example, real robot teleoperation + body-free collection, real robot teleoperation + simulation, body-free collection + simulation, and full-route coverage.

There are more players adopting cross-route solutions than those betting solely on any single route.

The industry often uses a "data pyramid" to describe the data structure required for robot training. Currently, no single data collection method can fully meet robot training needs on its own.

△

Reality 2: The largest number of players bet solely on the real robot teleoperation route

Let's analyze each technical route one by one.

The real robot teleoperation route has the most solo bettors: 22 companies, accounting for 31%.

Among these 22 players, 13 are state-owned data platforms, 7 are robot companies (that produce robot hardware or develop large embodied models), 1 transitioned from the AI data annotation industry, and 1 is a cross-over player from the industrial equipment manufacturing sector.

Robot companies naturally choose real robot teleoperation collection thanks to their hardware advantages and real-world demand.

State-owned data platforms, on the other hand, have the advantage of "not shying away from heavy investment". Teleoperation is a capital-intensive route that requires purchasing robot bodies, renting spaces, and hiring operators—all resources that state-owned platforms can easily mobilize.

15 companies bet solely on the body-free collection route, accounting for 21%.

These are the youngest companies in the track, with the vast majority founded after September 2024.

The body-free collection route also has the most diverse technologies, including subcategories like Ego (first-person) perspective, UMI, motion capture, sEMG electromyography, tactile collection, and more.

Only 2 players bet solely on simulation synthesis: Songying Technology and Motphys.

Previously well-known players in the simulation track have now chosen to "not put all their eggs in one basket".

For example, Guanglun Intelligence, which once centered its business on simulation data, has started collecting human demonstration data. Galaxy General, once one of the most staunch simulation advocates, launched a full-body teleoperation system in June this year, gaining teleoperation data collection capabilities.

There are two reasons for this shift: Externally, the supply of real robot data and human demonstration data is rapidly increasing while prices continue to drop, eroding the scale and cost advantages of simulation data. Internally, there's still no good solution to the "sim2real gap"—it's very difficult to faithfully reproduce friction, deformation, force feedback, and tactile sensations from the real world.

Only 1 player bets solely on the internet video distillation route: Shutu Technology.

This company extracts multi-modal robot training data from monocular RGB videos on the internet, claiming it can reduce overall collection costs to 0.5% of the industry average.

Embodied Data Collection

Who are the players?

Reality 3: Independent data service providers have become the largest player group

If we categorize players by their identity instead of technical routes, the 97 players fall into 5 categories:

39 independent data service providers, accounting for 40%;

25 state-owned data platforms, accounting for 26%;

24 robot companies, accounting for 25%;

5 cross-over companies from industrial and IT fields, accounting for 5%—for example, companies from logistics, equipment manufacturing, automation engineering, and other sectors;

4 large tech platform companies, accounting for 4%—such as Huawei and JD.com.

As we can see, the largest player group is independent data service providers.

This shows that the embodied data sector has grown into an independent track, no longer just an auxiliary department of robot companies.

△

Reality 4: Two-thirds of players are "embodied-native", one-third are "cross-over transformers"

Let's try another classification method, dividing all embodied data industry players into two groups: "embodied-native" and "cross-over transformed".

"Embodied-native" companies were founded with their core business focused on embodied data or embodied intelligence; "cross-over transformed" companies mostly transitioned from AI data annotation, autonomous driving, motion capture, or industrial sectors.

Among the 97 players, 65 are "embodied-native", accounting for 67%; 32 are "cross-over transformed", accounting for 33%.

Breaking it down further, the composition of data collection companies and data infrastructure companies is completely reversed.

Of the 70 data collection companies, 57 are "embodied-native", accounting for roughly 80%; of the 27 data infrastructure companies, 19 are "cross-over transformed", accounting for roughly 70%.

Why does infrastructure attract transformed companies, while the collection sector is mostly new players?

Many embodied data infrastructure players are AI data annotation companies, such as Hightech, Datatang, and Testin. Their accumulated pipelines, quality control, and delivery capabilities are very suitable to be directly transferred to embodied data infrastructure work.

Since there's no ready-made embodied data, the collection segment requires building assets from scratch—old players have no inherent advantages, while new companies can move forward lightly.

Embodied Data Collection

Production Capacity and Layout

Reality 5: The industry's total annual production capacity is 1.6 million to 1.8 million hours, with a short-term goal of 15 to 20 times expansion

What's the current embodied data production capacity? How big is the gap from market demand?

According to Qbit's incomplete survey, the current annual production capacity of the embodied data industry is: 1.6 million to 1.8 million hours + 70 million to 80 million data entries.

The industry's short-term goal is: In the next 1 to 3 years, produce 25 million to 35 million hours of data + data at the 100-million-entry level. If we only look at the hour count, the short-term goal is 15-20 times the current capacity.

It's worth noting that since different institutions use different reporting standards, there is no unified conversion rate between hours and data entries, so both metrics are listed here.

These numbers only count real robot teleoperation data and body-free collection data, excluding simulation synthesis data. The capacity is conservatively estimated based on publicly disclosed information from companies/platforms—the actual number may be even higher.

The total demand for robot training data is still unknown. But we can reference the large language model (LLM) benchmark: LLMs can consume all the ready-made text corpus on the internet, while robots can only collect their required data one piece at a time. Statistics show that as of the start of this year, the total amount of high-quality real physical interaction data worldwide is only about 500,000 hours—less than 1/20,000 of the training data volume for LLMs.

From another perspective, even if the short-term capacity goals are fully achieved, compared to the data volume of LLMs, it may only be enough to reach the starting line. There is still a huge gap between production capacity and demand.

Embodied Data Collection

Reality 6: 60% of China's provinces have built data collection factories, with the Yangtze River Delta having the largest distribution currently

Where is all this data collected?

According to Qbit's incomplete survey, data collection factories have been established in 20 provinces across China, among which state-owned-backed data collection factories cover 16 provinces.

Data collection factories are mainly distributed in the Yangtze River Delta, Beijing-Tianjin-Hebei region, and Pearl River Delta. The Yangtze River Delta ranks first with 30 factories.

Many third- and fourth-tier cities with lower labor costs have also become locations for data collection factories, such as Suqian, Zigong, Chenzhou, Yuncheng, and Deqing.

The distribution pattern is related to technical routes: teleoperation data collection factories are spread across various provinces, while light-asset body-free route companies cluster in first-tier cities.

Many cities are building city brands focused on being data collection hubs.

For example, Wuxi is the first city in China to propose the concept of city-wide full-domain data collection. Its most important initiative is to encourage manufacturing and service industry enterprises to open up their production lines and platforms, using real scenarios as data collection sites to gather the most scarce and practical data for robots.

△

Where is the money flowing?

Reality 7: 15 independent embodied data service providers raised about 4.47 billion yuan in the past year

Let's look at the most telling indicator: capital.

Since the data business of robot companies cannot be separated from their overall financing, we've identified 15 "independent embodied data service providers" with financing records in the past year—their financing situations are highly representative of the industry.

To clarify, there are three selection criteria for "independent embodied data service providers": they don't produce general-purpose robot bodies, they don't train embodied models, and their core business is embodied data.

According to Qbit's incomplete survey, in the past year (July 1, 2025 to July 1, 2026), these 15 "independent embodied data service providers" completed 34 financing rounds, totaling about 4.47 billion yuan.

The financing periods are highly concentrated. Between April and June 2026, over 40% of the total financing events took place. This is closely related to the capital frenzy across the entire embodied intelligence industry in the first half of this year.

Qbit previously calculated that in the first half of 2026, the entire embodied intelligence industry raised about 4