Serving as a "teacher" for humanoid robots, it underpins a 100-billion market.
Before robots can truly enter factories and households, someone must first teach them how to "work like a human."
You might have come across a video on social platforms showing workers on Indian assembly lines sorting, assembling, sewing, and cutting as usual, while cameras mounted on their heads and wrists record every single movement detail. This is essentially data collection for training humanoid robots.
In China, similar work has started to permeate the gig economy. Recruitment posts for "embodied intelligence data collectors" have emerged densely, with the slogans "daily payment, work from home, no academic requirements" attracting a large number of job seekers. Some people repeatedly grab water cups, organize clothes, and move items, serving as the "AI trainers" for robots.
Behind this lies the data hunger that the embodied intelligence industry is currently facing. For humanoid robots to transition from demonstration videos to real-world scenarios, they cannot rely solely on models and hardware entities—they also require massive volumes of authentic, clean, and reusable motion data.
The three collection paths—entity-free first-person crowdsourcing, real-robot teleoperation, and simulation generation—have each staked their positions, and a competition centered on data quality, cost, and efficiency is unfolding.
Cyber Assembly Lines
Both the physical bodies and the "brains" of humanoid robots are evolving, but the more they enter practical scenarios like factories, warehouses, and households, the more specific the problems become:
Asking a robot to fetch a glass of water placed in the center of a 1.2-meter-high dining table, which is a wide-mouthed handled mug surrounded by other clutter, requires the robot to judge where to extend its hand and how to grasp the cup.
These capabilities do not emerge out of thin air. They require a large volume of authentic, continuous, and annotatable data for training.
After the "data desert" gradually became a consensus across the industry, the upstream track of embodied intelligence data collection began to heat up rapidly. According to the "2026 Research Report on the Layout of the Embodied Intelligent Robot Data Industry" jointly released by Zoysia Automotive Research and the Shuimu Tsinghua Research Center, the global embodied intelligence data market size reached 242 million U.S. dollars in 2025, representing a year-on-year increase of 181.4%; among this, the Chinese market hit 5 billion RMB, a 203% year-on-year surge, accounting for approximately 40% of the global market.
The report forecasts that the global market will register a compound annual growth rate (CAGR) of 85.0% from 2025 to 2030, with the total size climbing to 5.25 billion U.S. dollars by 2030.
Against this backdrop, diverse players are flocking into the field. Third-party data service providers have entered the market—for instance, LightWheel Intelligence focuses on a hybrid simulation + entity-free solution, securing 2 billion RMB in financing within half a year, reaching a valuation exceeding 10 billion RMB, and serving multiple leading enterprises. Some humanoid robot manufacturers have opted to build large in-house collection bases, with Agibot even opening the AGIBOT WORLD 2026 dataset covering full-spectrum embodied intelligence research. Internet giant JD also announced plans to construct the world's largest embodied intelligence data collection center.
Entity-free Ego (first-person wearable) collection is currently the most popular path in the industry. It does not require a physical robot body, as data can be generated simply by having users wear cameras to record human movements, making it the track with the lowest entry barriers and the largest number of participants.
The Cervo team (Henan Nasixiongde Network Technology Co., Ltd.), which focuses on global real-scenario collection, is one of the players on this path, with business covering China, India, and South America. Ray, a partner of the team, introduced that all data collected by the company is sourced from real convenience stores, factories, and households. He believes that recording authentic operating habits delivers better adaptability for robot deployment in real scenarios.
A complete collection workflow starts with demand alignment: the client clarifies scenario types, hardware specifications, collection duration, and regional requirements, the team connects with offline partner scenarios, trains frontline operators on equipment usage and motion standards, and after on-site recording, the data is aggregated to the cloud via storage cards before being sent to partner teams for cleaning and annotation.
There are two tiers in hardware solutions: entry-level setups use GoPro action cameras or head-mounted mobile phones, which are low-cost but have limited field of view, with hands easily moving out of the frame, resulting in an effective data rate of roughly 93% to 95%—requiring strict training for operators on head movement. High-end solutions adopt six-eye panoramic collection devices, costing 3,000 to 8,000 RMB each, which can cover a 300-degree field of view with an effective data rate close to 100%, making them the preferred choice for well-funded leading enterprises.
In the Chinese market, entity-free collection has already penetrated the mass gig economy. An embodied intelligence data collector told "Baobian" that a part-time position at a company in the Pearl River Delta offers a daily wage of around 150 RMB, requiring 10 hours of work per day with three cameras worn on the head, hands, and other parts, performing standardized movements like making fists and storing items repeatedly. "The work is like a cyber assembly line—you repeat hundreds or thousands of times, your hands and neck get sore, the equipment often malfunctions, and you might get fined if you fail the probation period."
At the same time, some companies are building in-house data collection teams. Zhou Xinghao, Business Director of Anhui Shudian Information Technology Co., Ltd., stated that the company has a total of over 100 full-time employees across the annotation and motion collection teams, mostly aged 20 to 26, with 45% holding bachelor's degrees, and monthly salaries ranging from 5,000 to 7,000 RMB. Recruitment prioritizes hand and limb coordination, patience, and willingness to learn proactively.
A data collector based in a first-tier city also mentioned that full-time positions offer higher pay, with a daily wage of around 250 RMB. The job requires on-site presence in fixed scenarios, delivering no less than 3 hours of valid data per day, with performance bonuses for extra output and wage deductions for insufficient work. Data collection supervisors oversee the process on-site, and employees work in pairs to cross-verify each other's work.
Costs are even lower overseas. Ray noted that in regions like Africa and Latin America, factory workers earn only 1.5 U.S. dollars per hour (equivalent to roughly 10 RMB per hour) for performing data collection tasks.
Low barriers, low costs, and rapid scalability are the core advantages of the entity-free path, but its drawbacks are equally prominent: data quality varies drastically, with a high proportion of invalid samples. Issues like non-standard movements, staged scenarios, and hands moving out of the frame are widespread, resulting in the actual usable rate of crowdsourced data across the industry generally falling below 30%—and the definition of "valid data" is largely controlled by clients.
In contrast to the "mass labor tactics" favored by the industrial sector, academic circles and technology-focused startups prefer the real-robot teleoperation and simulation collection paths, prioritizing data accuracy and reproducibility.
An industry insider who previously worked as a research assistant in a university AI lab and now works at an embodied intelligence startup explained that real-robot teleoperation collection involves remotely controlling robotic arms or humanoid robots to complete tasks, while synchronously recording multi-modal time-series data such as joint angles, force feedback, and visual footage. "This data fully aligns with the motion logic of the robot body, with no cross-hardware deviations, making it the core data source for the model fine-tuning stage."
However, real-robot collection has extremely high barriers: a single humanoid robot body costs hundreds of thousands of RMB, strict requirements are imposed on the data collector's hand-eye coordination, and the core challenge of multi-sensor time-series alignment must be addressed. Different devices have varying frame rates and latencies, and even millisecond-level time misalignments can drastically reduce data value.
The insider's team solved this problem by increasing the collection frame rate and customizing hardware with unified timestamp interfaces. The entire pipeline requires collaboration between software and hardware teams, making rapid large-scale expansion impossible, unlike with entity-free collection.
Simulation collection is another complementary path, which generates motion samples in bulk in virtual environments, eliminating hardware and venue costs.
But the insider admitted that there is an insurmountable Sim-to-Real gap between simulation and the real world—factors like object friction, deformation, and lighting differ from reality. Models trained purely on simulation data perform poorly when deployed on physical robots, and can only serve as supplementary samples during the pre-training stage, unable to support high-precision tasks.
How to Collect Better Data?
In practical deployment, none of the three data acquisition paths are inherently superior. The path choices made by different teams essentially boil down to striking a balance between cost, efficiency, and data quality.
The core reason for the rapid explosion of entity-free collection lies in its extremely asset-light model.
Ray's team has only 8 core members across China and the U.S., with no full-time collectors or annotators on the payroll. The core team only handles client engagement, scenario resource matching, and project coordination. Equipment can be rented, labor is paid by the hour, and venues leverage existing physical resources—no heavy investment in robot bodies is required. The track can be entered with just tens of thousands of RMB in startup capital, which is the core reason why a large number of small and medium-sized teams have flooded in.
However, data quality issues are unavoidable, the most typical of which is "spurious correlation."
Zhou Xinghao cited a classic example: a collector is accustomed to blowing on a teacup to cool it down before picking it up, leading the model to identify "blowing" as a necessary precondition for grasping a cup. Since most humanoid robots do not have a "blowing" function, the trained robot will pause in front of an empty cup every time it tries to grasp it, completely deviating from real-world usage logic.
"More data does not necessarily mean better, but dirtier data definitely means worse," Zhou Xinghao summarized. In his view, the negative impact of low-quality data far exceeds that of insufficient data. "Insufficient data can be supplemented with additional collection, but contaminated data requires cleaning and rework, which is essentially taking a step backward."
During the industry's early wild growth phase, a large number of intermediaries purchased low-quality data at low prices, repackaged it simply, and resold it. This "garbage data" circulated repeatedly across the industry, further exacerbating the "data desert"—despite the seemingly massive total data volume, there are very few high-quality samples that are truly usable.
In contrast, the real-robot path offers controllable quality but is constrained by high costs, making large-scale deployment difficult.
A single humanoid robot body costs hundreds of thousands of RMB, not to mention the supporting venues, maintenance personnel, and calibration equipment. The investment required for a single collection pipeline is dozens of times that of the entity-free solution. Furthermore, different robot brands have varying motion spaces and hardware parameters, resulting in extremely low data reusability across different bodies. Replacing hardware almost requires re-collecting the entire dataset, leading to a far lower return on investment than the entity-free approach.
The simulation path, on the other hand, is limited by technological maturity—the precision of physics engines is insufficient to support complex interactive tasks, so it cannot become the mainstay of data supply in the short term.
A truly mature collection solution rarely bets on a single path, but instead combines them in layers based on specific requirements.
Leading internet giants and top humanoid robot manufacturers adopt a combined "80/20 ratio" strategy: 80% low-cost entity-free data is used for model pre-training, expanding the data volume to enhance basic generalization capabilities; 20% high-precision data collected from real robots is used for scenario fine-tuning, ensuring the success rate of deployed tasks.
Zhou Xinghao summarized this as "rough data lays the foundation, and precise data enables optimization," noting that the two types of data correspond to different stages of model training and are both indispensable.
Small and medium-sized startups with limited budgets do not have such conditions, and most can only purchase low-cost crowdsourced data to support basic R&D.
Universities and research institutions face a similarly awkward situation. In academic circles, students mostly build collection pipelines independently, which are small in scale, lack standardized quality control, and produce unstable data quality—though the upside is that costs are low, and teams can flexibly customize niche tasks to adapt to cutting-edge algorithm research.
Between quality and scale, the industry is searching for a middle ground—for example, using more standardized management to increase the usable rate of low-cost data.
Zhou Xinghao's team underwent a transformation from pure crowdsourcing to on-site collection. Building on this, the team established a three-tier quality control mechanism, and attached a "data ID card" containing collection information and quality inspection records to each piece of data, enabling full traceability of issues.
After implementing this system, the team's data usable rate rose from the industry average of 30% to 50% to around 80%.
This also demonstrates that embodied intelligence data collection is evolving from its early wild growth phase to a more refined production management model.
From Selling Data to Selling Skills
Beyond technological integration, the business model of the embodied intelligence data collection track is also upgrading.
The track that solely sells raw data is falling into a vicious cycle of low-price competition. Leading data service providers have begun to explore extending upstream, shifting from "selling data" to "selling skills."
Ray's team is planning a transformation path toward Skill-as-a-Service: leveraging the massive volume of scenario data it has accumulated, fine-tuning dedicated skill models for individual industrial processes based on open-source foundational models, and pairing them with robotic arms to form complete solutions.
The core logic of this model is cost substitution: labor costs in European and American factories are around 30 U.S. dollars per hour, while the comprehensive operating cost of a robotic arm paired with a customized skill model can be controlled between 13 and 15 U.S. dollars per hour, directly reducing labor costs. Service providers charge skill service fees based on working hours, delivering far greater profit margins than simply selling data.
"In the future, we will not be data suppliers, but robot skill providers," Ray described.
When discussing the long-term development of the industry, the three interviewees reached a highly consistent judgment: the industry has not yet witnessed a decisive "GPT moment," but it is certain that teams that rely on demonstration videos to secure financing and hype concepts will be eliminated. Only teams that possess real offline scenario resources and full-link quality control capabilities will survive.
In terms of deployment pace, industrial semi-structured scenarios will be the first to achieve commercial breakthroughs—a consensus shared across the entire industry.
Zhou Xinghao analyzed that industrial scenarios offer controllable environments, clear task boundaries, and quantifiable economic value. The payback period for replacing a single workstation can be precisely calculated, giving enterprises strong willingness to pay. The three processes of assembly, sorting, and quality inspection will be the first to achieve full commercial viability.
The aforementioned insider at the embodied intelligence startup also believes that research in both academic and industrial circles is shifting toward industrial scenarios, with model precision in single fixed scenarios already approaching the deployment threshold.
As for the general-purpose household humanoid robots that the public anticipates, all parties agree that there is still a long journey ahead.
Household scenarios are open and random, with infinite interfering factors, requiring an exponentially growing volume of data. Coupled with persistently high hardware costs and low consumer willingness to pay, widespread adoption in the short term is nearly impossible.
In the next 3 to 5 years, household scenarios will first see the emergence of dedicated intelligent devices like kitchen robotic arms and clothes-folding machines, rather than all-around humanoid robots. The deployment of general-purpose humanoid robots will require at least 10 to 20 years of technological and data accumulation.
From assembly line workers in Indian factories to part-time collectors in China, the "mass labor tactic" is rapidly filling the data gap in embodied intelligence, while the "meticulous, deliberate work" of real-robot collection and simulation is safeguarding the bottom line of data quality. The two paths have not yet converged, but heterogeneous hybrid training has already built a bridge connecting them.
It is certain that the wild growth phase will eventually come to an end. Establishing unified collection standards and generating high-quality, traceable data is the long-term survival rule for this track. Only when enough deep wells are dug in the "data desert" will the commercial deployment of embodied intelligence truly reach its inflection point.