A New "Winning Move" for Physics AI—"Data Infrastructure" Stands on the Eve of Explosion

```

The focus of competition in the humanoid robot industry is quietly shifting.

On July 15, Guojin Securities released an in-depth research report pointing out that as hardware matures, the key variable determining the upper limit of embodied intelligence has shifted from mechanical performance to high-quality data. Compared to computing power and hardware, data has become the scarcest production factor in the era of physical AI.

The report estimates that to train embodied intelligence models with practical application capabilities, the industry needs at least 10 million hours of multimodal interactive data, while the globally accumulated data volume is still less than 5% of the demand. This huge supply-demand gap is driving a new round of construction for data collection centers, data services, and collection equipment.

Guojin Securities believes that the data collection industry chain will be the first to benefit. Among them, core equipment suppliers such as cameras, IMUs, tactile sensors, data service providers with advantages in cost and quality, and enterprises possessing closed vertical scenario data resources are expected to become the most certain benefiting directions in the industry boom. Data infrastructure is taking center stage in the era of physical AI.

Hardware enters a convergence period, robot competition shifts to "brains"

The report points out that the humanoid robot industry has entered a new stage of development. After years of rapid iteration, hardware solutions are gradually maturing, and the industry's focus is shifting to intelligent model training capabilities, with high-quality data being the core determining factor for model performance.

This trend has also received policy support. In November 2023, the Ministry of Industry and Information Technology issued the "Guidelines for Innovative Development of Humanoid Robots," explicitly proposing to build a robot training database and continuously expand high-quality, multimodal data, providing fundamental support for the training of the robot "brain."

The importance of data has already begun to be reflected in industry orders. According to Xinghe Frequency statistics, among humanoid robot orders exceeding 10 million yuan in 2025, there are 15 data collection projects, accounting for 31.25%, surpassing industrial manufacturing (27.08%) and commercial services (10.42%) for the first time, becoming the largest segment by proportion.

Guojin Securities believes that this means data has evolved from a supporting research and development link to an independent industry track. At the same time, the physical AI market has also entered a period of rapid expansion. Future Markets predicts that the global physical AI market size will grow from $383 billion in 2026 to $3.26 trillion in 2040, maintaining a high compound growth rate over the next decade or more, with data infrastructure construction expected to usher in an investment peak first.

Nationwide construction of data collection centers accelerates, Beijing and Yangtze River Delta take the lead

With rapidly growing industry demand, construction of national data collection centers has clearly accelerated. Statistics from GG Humanoid Robot show that as of now, there have been at least 15 data collection and training factories built or under construction nationwide, forming an industrial pattern of "Beijing leading, Yangtze River Delta clustering, and rapid layout in central and western regions."

The main data factories already in operation are expanding:

Zhiyuan Data Collection Factory (to be put into use in September 2024), with an area over 3,000 square meters, can collect 30,000-50,000 data items per day, covering five major categories including home, catering, industry, and over 200 sub-scenarios.
Tianjin Pasini Embodied Intelligence Super Data Factory (to be put into use June 2025), with an area of nearly 12,000 square meters, a single day collection capacity of 50,000, relies on real human hand motion capture, annual collection capacity nearing 200 million, efficiency 3 to 6 times higher than traditional teleoperation.
Wuxi Embodied Intelligence Robot Industrial Data Collection and Training Center, with an area of about 7,000 square meters, designed for annual output of over 10 million industrial data items.

Robot body manufacturers are also beginning to lay out data capabilities on a large scale. Leju, UBTECH, and Zhiyuan have all built data collection systems; Pasini Perception is planning to build five data factories; UBTECH's Walker S series has become the most widely used data collection robot, winning orders for data collection projects worth over 750 million yuan, and is expected to surpass 800 million yuan in annual orders in 2025.

Three technical routes run in parallel, first-person data becomes new hotspot

The report believes that the industry currently mainly adopts three data collection routes: teleoperation, motion capture, and first-person (EGO), each balancing cost, accuracy, and scalability.

Teleoperation remains the most mature real data collection method, including posture control, visual control, and opto-inertial fusion, enabling millimeter-level precision and data collection frequency approaching kilohertz, but equipment and labor costs are high, so the industry is moving towards lighter and lower-cost solutions.

Motion capture mainly includes inertial and optical motion capture. Among these, Noitom's inertial motion capture products hold about 70% of the global market share, already providing data collection services to more than 60 robot companies; optical capture offers higher precision—Leyard's OptiTrack can achieve sub-millimeter positioning—but still needs to solve the occlusion problem, and the industry is currently improving stability through opto-inertial fusion, local magnetic fields, and AI calculation solutions, etc.

First-person (EGO) data collection is the fastest-growing direction. The report notes NVIDIA's GEAR Lab will launch its EgoScale framework in 2026; after pre-training with 20,854 hours of first-person video, the 22-DOF dexterous hand's manipulation success rate increased by 54%, demonstrating the value of first-person data for embodied intelligence training and rapidly driving demand for relevant data collection equipment.

At the same time, simulation data has also become an important supplement. Yinhang Universal has established an automated simulation data generation process, able to generate 1 billion frame-level data sets within a week, making training about 1000 times more efficient than traditional solutions. However, the reliability of sim-to-real transfer in the real world remains to be further validated.

Three major investment directions emerge: equipment, data services, and vertical scenarios

The report points out that around the data collection industry chain, equipment, data services, and vertical scenarios are the three most favored directions. On the equipment side, cameras, IMUs, and tactile sensors form the core entry point for data production. Among these, non-physical data (first-person videos and general operation interface data) is expected to account for 40%–50% of total data time, while real machine teleoperation accounts for about 30%, driving EGO cameras and UMI grippers to become the fastest-growing data collection device types.

Data service is still in the early stage of competition. Leading tech companies plan to collect over 10 million hours of real-world video data in two years, mobilizing hundreds of thousands of people to participate in data production; while some startups have proposed longer-term goals of collecting billions of hours. Among current industry participants, some companies have joined ultra-large-scale real data plans and signed with embodied intelligence companies, while others have accumulated tens of thousands to hundreds of thousands of hours of basic data in EGO data collection and dexterous hand teleoperation data.

Compared to general-purpose robots, commercialization of vertical scenarios may progress faster. Due to relatively closed industry scenarios and smaller data demands, fields such as automobile manufacturing and power inspection are likely to achieve data closed loops first. For example, some companies use automotive and lithium battery industrial scenarios to build embodied intelligence industrial data centers, with data covering real production lines of mainstream car companies and electronics manufacturers; others have long accumulated multimodal data in specialized scenarios such as power inspection, forming high barriers in segmented fields.

Risk Disclosures and DisclaimerThe market carries risk, and investments should be approached cautiously. This article does not constitute personal investment advice and does not take into account any particular user's specific investment objectives, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their circumstances. Investment decisions based on this content are made at one's own risk. ```