In 2026, the center of gravity in embodied intelligence is shifting. The industry is moving from the challenge of building robots to the challenge of feeding those robots enough data. At corporate launches, industry forums, and investor briefings, data has become the most frequently discussed topic. The language of the sector has changed. Hardware demonstrations still matter, but they no longer dominate the agenda. The decisive question is whether embodied intelligence systems can obtain enough physical-world experience to move from controlled pilots to general capability.
Yao Maoqing, chairman and CEO of Mifeng Technology, described the bottleneck directly: the biggest obstacle and challenge for physical AI has not yet reached the stages of computing power and algorithms. Compared with large language models, physical AI remains at an earlier stage, and the data problem is becoming more prominent. That assessment captures the mood across the embodied intelligence industry. The next phase of competition will not be decided only by robot bodies, actuators, or foundation model architectures. It will be decided by the quantity, quality, diversity, and deliverability of data.
Several developments illustrate the scale of the infrastructure wave. Mifeng MEgo, associated with the Zhiyuan ecosystem, has put 20,000 sets of equipment into production and formally produced 1 million hours of body-less data. Qianxun Intelligence has built a distributed large-scale data collection network with more than 400,000 real collection points across the country. JD.com has proposed building the world’s largest data collection system. According to IT Juzi, 25 Chinese embodied intelligence data startups raised more than RMB 17 billion in the first half of 2026. These figures do not merely describe a boom. They describe the emergence of a new layer in the embodied intelligence value chain.

As millions of hours of data begin to enter the market, the real question is no longer whether data exists. The real question is whether the data can be used. Data standards, quality evaluation, and privacy compliance are becoming the new dividing lines. In embodied intelligence, the hidden variables are becoming as important as the visible hardware. The industry is discovering that data is not a commodity simply because it has been collected. It becomes a commodity only when it is standardized, validated, and accepted by customers.
1. A Structural Shift: From Building Robots to Feeding Embodied Intelligence
The shift now underway is structural. For years, the embodied intelligence narrative focused on embodiment itself: humanoid form, dexterous hands, locomotion, battery life, and hardware cost. Those issues remain important. Yet the center of gravity has moved toward the data layer that allows embodied intelligence models to learn. Without enough physical interaction data, even advanced robot bodies remain limited to narrow tasks. Without diverse and high-quality data, embodied intelligence cannot develop the physical common sense that humans acquire through daily experience.
The comparison with large language models is now common in the industry. A large language model can be compared to a scholar who has read widely. An embodied intelligence model is more like an infant who must learn from scratch how to interact with the world. The scholar can learn from all the text humans have written. The infant must touch fire to know heat, and must pick up a cup to understand the mechanics of grasping. This difference is not a minor technical detail. It creates a data discontinuity that shapes the entire embodied intelligence industry.
That discontinuity is visible in the demand structure. According to data cited in the industry, in 2025 data collection demand accounted for 31 percent of ten-million-level humanoid robot orders. In 2026, that proportion is expected to exceed 50 percent, making data collection the largest source of demand for humanoid robots. The implication is significant. The robot itself is no longer the only product being sold. The data pipeline, the collection device, the annotation process, the quality report, and the delivery contract are becoming products in their own right. Embodied intelligence is creating a market for data infrastructure before it has created a mass market for general-purpose robots.
Body and model manufacturers are one group of players. They rely on their own robots and core sensing hardware to build data collection systems. After satisfying internal needs, they sell standardized datasets externally. Mifeng Technology, incubated by Zhiyuan Robotics, completed three financing rounds in only half a year. Its self-developed MEgo series of body-less collection devices has been mass-produced and delivered in more than 20,000 sets. Qianxun Intelligence independently developed wearable data collection devices and has more than 400,000 real collection points in more than 100 cities. These examples show how quickly embodied intelligence companies are moving into the data business.
On the other side, a group of independent shovel sellers is rising rapidly. Guanglun Intelligence, Jianzhi Robotics, Noitom Robotics, and others form the first tier of domestic data infrastructure. Their emergence signals a change in industrial division of labor. Before 2024, almost all data work was done internally by robot companies. By 2026, more and more small and medium-sized body manufacturers are choosing to buy datasets and collection services rather than build expensive internal data teams. The value chain is migrating from hardware-body-centric to data-infrastructure-driven. Data itself is becoming a tradable, deliverable, standardized commodity.
| Indicator | Reported Figure or Status | Implication for Embodied Intelligence |
|---|---|---|
| Data collection share of ten-million-level humanoid robot orders in 2025 | 31 percent | Data demand already formed a major portion of robot orders. |
| Expected data collection share in 2026 | More than 50 percent | Data collection becomes the largest demand source for humanoid robots. |
| Global available high-quality real data | Hundreds of thousands to the million-hour level | The supply base remains far below the scale needed for general capability. |
| Data required for an embodied foundation model ChatGPT moment | At least tens of millions of hours | The gap exceeds 99 percent. |
| Chinese embodied intelligence data startup financing in the first half of 2026 | 25 companies raised more than RMB 17 billion | Capital is flowing into the data layer of embodied intelligence. |
| Mifeng MEgo production and data output | More than 20,000 sets of equipment; 1 million hours of body-less data | Body-less collection is scaling into real-world environments. |
| Qianxun Intelligence collection network | More than 100 cities; more than 400,000 real collection points | Distributed collection networks are expanding geographic and scene diversity. |
| QuantumBit count of data collection companies | 70 companies; 30 pursue multiple routes, or 43 percent | No single collection method is sufficient for embodied intelligence training. |
2. The 99 Percent Data Gap and the Rise of a New Industry
The scale of the gap explains why the data infrastructure wave has become so intense. The “Embodied Intelligence Training Ground Research Report (2026)” released by the China Academy of Information and Communications Technology shows that the world’s currently available high-quality real data is only at the level of hundreds of thousands to one million hours. An embodied foundation model would need at least tens of millions of hours of data to reach a ChatGPT moment. The gap exceeds 99 percent. The industry is far from the critical point needed to support capability emergence.
This gap is not a temporary shortage. It is a structural feature of embodied intelligence. Text and images can be scraped from the internet at massive scale. Physical interaction data must be collected in the real world, with sensors, devices, operators, and scenes. It must capture contact, force, motion, failure, recovery, and context. It must be labeled, synchronized, validated, and stored. Each step introduces cost, complexity, and potential error. The data problem therefore cannot be solved by simply waiting for larger models or faster chips. It requires an industrial system.
That system is now forming. The first tier includes body and model manufacturers that collect data for their own training and then sell standardized datasets. The second tier includes independent data service companies that focus on collection hardware, crowdsourcing operations, annotation, quality control, and delivery. The third tier includes platforms and marketplaces that aggregate demand and supply. The boundaries between these tiers are still shifting. Some robot companies are becoming data vendors. Some data vendors are moving into model evaluation. Some platforms are building their own collection networks. In embodied intelligence, vertical integration and horizontal specialization are occurring at the same time.
The rise of independent service providers marks a qualitative change in industrial division of labor. In the early phase, a robot company had to own the entire data pipeline because no external supplier understood its hardware, sensors, and model requirements. By 2026, external suppliers have accumulated enough domain knowledge to serve multiple customers. Small and medium-sized body manufacturers can now purchase datasets and collection services instead of building expensive data teams. This lowers the barrier to entry for robot development, but it also raises the bar for data suppliers. They must deliver datasets that are compatible with different embodiments, sensor suites, and task definitions.
Data is therefore becoming a product with specifications. Customers are no longer satisfied with a raw folder of video files. They want synchronized multimodal streams, consistent frame rates, known camera parameters, annotation schemas, task metadata, quality reports, and acceptance criteria. The embodied intelligence industry is learning what other data-intensive industries learned earlier: the value of data lies not in its volume alone, but in its fitness for use. A dataset that cannot be validated, reproduced, or integrated into a training pipeline has little commercial value, no matter how many hours it contains.
3. Four Parallel Technical Routes for Embodied Intelligence Data
Faced with the enormous data gap, the industry has not converged on a single technical route. Instead, it has differentiated into four mainstream approaches. Each route has distinct advantages, limitations, and roles in the embodied intelligence training pipeline. The coexistence of these routes is not a sign of confusion. It is a rational response to the fact that embodied intelligence requires different types of data at different stages.
- Real-machine teleoperation data collection. In this route, an operator remotely controls a physical robot body through a master device to complete tasks. The system directly records all sensor data from the robot itself. The advantage is that the data is fully aligned with the target embodiment. It is irreplaceable for post-training and real-machine deployment. The disadvantage is extremely high cost and slow expansion. To produce tens of thousands of hours of high-quality real-machine data, more than 1,000 robots may need to collect continuously for more than half a year. This route is essential for embodied intelligence deployment, but it cannot alone supply the scale needed for pretraining.
- Body-less data collection. This has been the hottest direction over the past year. Device forms include head-mounted cameras, wrist sensors, and tactile gloves. No physical humanoid robot body is required. The MEgo series from Mifeng belongs to this route. Its greatest advantage is that it can spread through crowdsourcing and enter a wide variety of real scenes. Mifeng has cumulatively produced more than 1 million hours of high-quality body-less data, all from real collection in open environments. Its collection devices have entered homes, offices, retail, production, and catering scenes. However, Yao Maoqing also acknowledges that body-less data cannot replace real-machine data. Body-less data is suitable for learning general representations during pretraining, while real-machine data is suitable for post-training. For deployment on a specific task, real-machine data for the corresponding body cannot be avoided.
- Simulation synthetic data. Simulation can generate extreme and long-tail scenarios without limit, with low marginal cost and convenient large-scale expansion. However, the Sim-to-Real gap persists. At present, simulation data accounts for only a small proportion of the training process and is used more for evaluation. Many practitioners have also observed that in some domestic scenarios, the per-hour cost of high-quality real-world data has already fallen below that of simulation. Simulation requires digital assets, and asset construction involves manual drawing, modeling, and other steps, and depends on GPU computing power for rendering, making overall costs high. Body-less data collection can dilute the cost of real data at a large scale.
- Internet video distillation. This route extracts action knowledge from existing video. Yao Maoqing positions it as a crutch for the current stage. When first-person perspective data is insufficient, it can only be used to obtain knowledge and representations. After body-less data reaches a scale of tens of millions of hours or more, this crutch can gradually be discarded. The route is useful as an interim supplement, but it does not solve the fundamental need for physical interaction data in embodied intelligence.
| Route | Mechanism | Main Advantage | Main Limitation | Primary Role in Embodied Intelligence |
|---|---|---|---|---|
| Real-machine teleoperation | Operator remotely controls a physical robot and records all onboard sensor data. | Data is fully aligned with the target embodiment; irreplaceable for post-training and real-machine deployment. | Extremely high cost and slow expansion; tens of thousands of hours may require more than 1,000 robots for over half a year. | Deployment, post-training, and task-specific real-machine validation. |
| Body-less data collection | Head cameras, wrist sensors, and tactile gloves collect data without a humanoid robot body. | Can be crowdsourced and scaled across real scenes; helps dilute the cost of real data. | Cannot replace real-machine data; better for pretraining than final deployment. | Pretraining and general representation learning for embodied intelligence. |
| Simulation synthetic data | Generate extreme and long-tail scenarios in virtual environments. | Unlimited generation and low marginal cost; convenient for large-scale expansion. | Sim-to-Real gap remains; small training share; mainly used for evaluation; digital asset and GPU rendering costs can be high. | Supplement for extreme, high-risk, and long-tail scenarios in embodied intelligence. |
| Internet video distillation | Extract action knowledge from existing video. | Useful when first-person data is insufficient; provides knowledge and representations. | Positioned as a current-stage crutch; can be discarded after body-less data reaches tens of millions of hours. | Interim supplement for embodied intelligence data needs. |
The four routes are currently running in parallel. In an incomplete count by QuantumBit of 70 data collection companies, 30 companies simultaneously pursue multiple collection routes, accounting for 43 percent. The reason is that no single data collection method can independently satisfy the training needs of robots. Body-less ego first-person data solves the scale problem of pretraining. Real-machine data guards the bottom line of real embodiment deployment. Simulation supplements extreme, high-risk, and long-tail scenarios. The industry judges that in the short term, there will not be a perfect technical route that wins across all scenarios.
This parallel structure has important implications for embodied intelligence strategy. A company that relies on only one data route may find its datasets incomplete for customer needs. A company that claims to replace all other routes with a single method may be overstating what its data can do. The most credible data suppliers are those that can explain which route they use, what role that route plays, what quality controls they apply, and how their datasets fit into a larger training curriculum. In embodied intelligence, data strategy is not a matter of choosing one route. It is a matter of orchestrating multiple routes across the lifecycle of model development.
4. The Coming Consolidation: From Fragmentation to Oligopoly
Beneath the surface excitement, a shakeout has already begun. Frontline practitioners judge that this track will eventually move toward an oligopoly. It will be difficult to maintain a fragmented, fully competitive state. Yao Maoqing told reporters that the embodied data industry will naturally move toward head concentration. This business simultaneously consumes engineering capability, operational capability, and financial strength. The requirements are not limited to collecting data. They include building hardware, managing crowds, ensuring consistency, processing multimodal streams, storing massive datasets, validating quality, and satisfying customer acceptance criteria.
Data collection is a heavy-asset game. Investment in tens of thousands of hardware sets is at the level of hundreds of millions of yuan. Storing PB to EB-level data and building data centers requires enormous additional expenditure. At the same time, it requires extremely strong engineering capability and large-scale crowdsourcing operational capability. Multiple thresholds overlap, and the survival space for small and medium-sized teams will continue to be squeezed. In embodied intelligence, the data layer may look like a software business from the outside, but its economics are increasingly infrastructure-like.
Customer requirements for data are also becoming higher. In the early stage of the industry, capital markets easily used cumulative hours of data as a core evaluation metric. After a round of practice, market cognition is iterating rapidly. One dimension is data quality. Yao Maoqing said that at the beginning of the year, companies were iterating, trial-and-error testing, and converging. Data collected with mobile phones or action cameras could still be sold. By the second half of the year, indicators such as 60FPS, 1080P, and stereo vision gradually became consensus. Demand is mainly defined by head customers.
Another dimension is data diversity. Yao Maoqing said that blindly collecting 1 million hours is not actually difficult. It can be achieved by repeatedly doing a few types of tasks. But such data has almost no value. The market has realized that when task scenes are single and data is homogeneous, the marginal benefit of continued expansion declines rapidly. No matter how good the model architecture is, without massive, diverse, high-quality interaction data from the physical world, physical cognition cannot be formed. This is a central lesson for embodied intelligence. Scale without diversity is not a solution. It is a cost.
Yao Maoqing proposed three indicators for evaluating the strength of a data collection enterprise. The first is the diversity of scenes and objects in the dataset. The second is whether the company can provide a detailed and verifiable data quality report. The third is how many datasets have completed customer acceptance and formed confirmed revenue, which is an objective proof of market competitiveness. These indicators shift the conversation from raw volume to deliverability. In embodied intelligence, a dataset that has passed customer acceptance is more meaningful than a dataset that merely exists.
| Dimension | Meaning for Embodied Intelligence Data |
|---|---|
| Completeness | The dataset must cover the required tasks, modalities, and context without missing critical information. |
| Consistency | Data streams, labels, and formats must remain coherent across collection sessions and devices. |
| Diversity | Scenes, objects, tasks, and conditions must vary enough to support general embodied intelligence. |
| Authenticity | Data must reflect real physical interaction rather than artificial or misleading approximations. |
| Usability | Datasets must be structured and documented so that they can be integrated into training pipelines. |
| Practicality | Data must serve real deployment needs and customer acceptance requirements. |
| Scalability | Collection and processing methods must support expansion toward tens of millions of hours. |
| Security | Privacy, compliance, and data protection must be addressed throughout the embodied intelligence data lifecycle. |
At the same time, the industry’s standardization process is accelerating. In August, the Ministry of Industry and Information Technology approved and released the “Artificial Intelligence Key Basic Technology: Quality Requirements and Evaluation Methods for Embodied Intelligence Datasets.” The standard will officially take effect on November 1, 2026. It covers eight dimensions: completeness, consistency, diversity, authenticity, usability, practicality, scalability, and security. The release of the standard marks a shift in embodied intelligence dataset construction from scale orientation to quality orientation.
For leading players, the standard is both a threshold and a moat. It is a threshold because companies that cannot meet the quality requirements may be excluded from serious procurement. It is a moat because companies that have already built quality-control systems, documentation, and acceptance workflows can use the standard to differentiate themselves. Yao Maoqing said that the competitive threshold for embodied data has shifted from whether one can collect to whether one can deliver stably. The industry will converge toward a small number of companies. In embodied intelligence, standardization is not the opposite of innovation. It is the condition for scaling innovation.
5. The Road to the ChatGPT Moment for Embodied Intelligence
The road to a true ChatGPT moment for embodied intelligence is still long. According to estimates, for an embodied foundation model to achieve a qualitative change, at least tens of millions of hours of real data are needed. This process still faces many challenges. The gap is not only a matter of collecting more hours. It is a matter of changing the operating model, the hardware ecosystem, the data supply chain, and the validation pipeline. The industry must move from project-based collection to industrialized data production.
Yao Maoqing said that the entire operating model needs a major upgrade. Linear replication through the existing model is extremely difficult. A way must be found to achieve a dimensionality reduction strike. He believes that linear expansion is too costly. The industry should rely on a crowdsourcing model to mobilize ordinary people with real skills from all walks of life to participate in data production. This vision is ambitious. It implies that embodied intelligence data collection could become a distributed economic activity, with contributors in homes, offices, stores, factories, restaurants, and other real environments. It also implies that the technical and operational systems must be robust enough to handle millions of contributors, thousands of device types, and many task categories.
At the same time, a series of engineering problems must be overcome. These include hardware consistency, multimodal alignment, automated data quality verification, and privacy compliance. Hardware consistency matters because data from different devices must be comparable. Multimodal alignment matters because vision, touch, force, motion, and audio streams must be synchronized. Automated quality verification matters because human review cannot scale to millions of hours. Privacy compliance matters because data collected in homes, offices, and public spaces may contain sensitive information. Each of these problems is difficult on its own. Together, they define the operational frontier of embodied intelligence data infrastructure.
The industry is also learning that data collection is not the end of the process. It is the beginning of a longer cycle that includes cleaning, annotation, validation, packaging, delivery, and feedback from model training. A dataset that is not used in training may still have cost but no return. A dataset that is used but performs poorly may damage the supplier’s reputation. A dataset that is accepted by a customer and integrated into a deployed embodied intelligence system becomes a reference case. This cycle is why confirmed revenue and customer acceptance are becoming important indicators. They prove that the data has crossed the boundary from raw material to usable infrastructure.
The competition among shovel sellers will therefore not be settled by one large contract or one impressive hour count. It will be settled by repeatability. Can the supplier collect the same quality across different cities and devices? Can it provide verifiable quality reports? Can it adapt to new tasks and new embodiments? Can it protect privacy and meet compliance requirements? Can it deliver on time and at scale? These questions are practical, but they determine which companies become the long-term infrastructure providers for embodied intelligence.
6. What the Shovel Sellers Must Prove
The 99 percent data gap is both an opportunity and a risk. It is an opportunity because the demand for embodied intelligence data is enormous and still growing. It is a risk because the cost of building an industrialized data pipeline is high, and the requirements are becoming stricter. Companies that treat data as a temporary service may find it difficult to survive. Companies that treat data as infrastructure may be able to build durable positions. The distinction matters because embodied intelligence will not advance on hardware alone. It will advance on the ability to feed models with the physical world.
The most important proof for a shovel seller is not a marketing claim about total hours. It is a verifiable record of datasets that have been accepted, used, and validated. That record must be supported by diversity across scenes and objects, detailed quality reports, and stable delivery. It must also be supported by a cost structure that can scale. The industry has seen that collecting 1 million hours is not enough if the data is repetitive. It has seen that body-less data can scale quickly but cannot replace real-machine data. It has seen that simulation can generate long-tail scenarios but still struggles with the Sim-to-Real gap. The winning data infrastructure provider will be the one that understands these boundaries and orchestrates them.
For the broader embodied intelligence ecosystem, the rise of data infrastructure has a clear implication. The companies that control high-quality, diverse, compliant, and deliverable physical interaction data will have a strong influence over which models can be trained and which robots can be deployed. This is why capital is flowing into the data layer. This is why robot makers are spinning out data companies. This is why independent data vendors are attracting attention. The shovel sellers are not peripheral to embodied intelligence. They are becoming central to it.
Yet the distance to the ChatGPT moment remains substantial. The world currently has high-quality real data at the level of hundreds of thousands to 1 million hours. Embodied intelligence needs at least tens of millions of hours. The gap exceeds 99 percent. Closing that gap will require more than investment. It will require standards, automation, crowdsourcing, privacy protection, multimodal engineering, and customer acceptance. It will require the industry to move from isolated collection projects to an industrialized data supply chain. The next phase of embodied intelligence will be defined by this struggle.
In that struggle, the most valuable companies may not be the ones with the most visible robots. They may be the ones that quietly collect, validate, and deliver the data that makes embodied intelligence possible. The shovel sellers of embodied intelligence are building the infrastructure for a world in which robots learn from physical experience. The 99 percent gap explains why they matter. The quality threshold explains why only a few will last. The road ahead explains why the work is far from finished.
7. The Operational Tests That Will Define the Next Phase
The next phase of embodied intelligence will be defined by operational tests rather than promotional milestones. The market has already moved past the stage where a large hour count alone can establish credibility. It is now asking whether data can be produced repeatedly, validated independently, and delivered under commercial terms. This shift is visible in customer behavior, investor questions, and the language used by leading suppliers. The data layer of embodied intelligence is becoming less speculative and more industrial.
- Scale with consistency. A supplier must be able to expand collection across cities, devices, and operators without creating incompatible datasets. In embodied intelligence, scale that produces inconsistent data is not scale at all. It is a future integration cost.
- Quality with verification. Quality must be demonstrated, not asserted. Detailed and verifiable quality reports are becoming part of the product. For embodied intelligence, a dataset without provenance and validation is difficult to use in serious training pipelines.
- Diversity with relevance. Diversity is valuable only when it matches the tasks and environments that embodied intelligence systems must handle. Scenes and objects must vary, but they must also be relevant to deployment. Random diversity can be as wasteful as repetition.
- Compliance with utility. Privacy and compliance cannot be treated as afterthoughts. They must be designed into collection and processing while still preserving the utility of the data. For embodied intelligence, data collected in real homes, offices, and public spaces carries both value and responsibility.
- Delivery with acceptance. The ultimate test is customer acceptance. A dataset that has completed acceptance and formed confirmed revenue is evidence that the supplier can meet commercial requirements. In embodied intelligence, acceptance is the bridge between data collection and data value.
- Cost with scalability. The cost structure must improve or at least remain sustainable as volume grows. Body-less collection can dilute real-data costs, but it cannot eliminate the need for real-machine data. Simulation can reduce marginal costs, but it cannot fully close the Sim-to-Real gap. The winning model will combine routes rather than rely on one.
These operational tests explain why the embodied intelligence data industry is likely to consolidate. They require capital, engineering discipline, operational know-how, and customer trust. They also explain why the 99 percent gap will not be closed by a single breakthrough. It will be closed by many parallel improvements across devices, standards, automation, crowdsourcing, and model feedback. The companies that can manage this complexity will become the long-term shovel sellers of embodied intelligence.
The final measure of success will be whether embodied intelligence systems can learn from the physical world with the same ease that language models learn from text. That goal remains distant. But the infrastructure being built today is the necessary foundation. The data gap is the defining constraint of the current era of embodied intelligence. The shovel sellers are the ones trying to remove it. Whether they succeed will determine how quickly embodied intelligence moves from demonstration to daily reality.
