Guojin Securities: Digital acquisition is the biggest shortcoming in physical AI, data infrastructure is about to explode

The Zhitong Finance App learned that Guojin Securities released a research report saying that data is the biggest sticking point on the physical AI circuit represented by humanoid robots, and the explosion of data infrastructure in the supply chain will benefit deeply. Data has become a key bottleneck limiting the development of physical AI, and the accelerated construction of data infrastructure is expected to drive the continuous expansion of the digital mining industry chain. (1) Digital acquisition equipment and supply chain: Focus on high ASP, well-structured racetracks such as cameras, IMUs, tactile sensors, etc.; (2) digital acquisition companies with barriers to data cost and quality; (3) vertical applications: focus on companies with resource barriers to segmented racetrack data scenarios.

Guojin Securities's main views are as follows:

Physical AI represented by humanoid robots, data is the most stringent field

Currently, hardware solutions in the humanoid robot industry tend to converge. “brain” training has become the decisive key, and the amount and quality of data directly determine the ability to generalize models. Real data on high-quality, high-fidelity physical interactions of robots is extremely lacking. The big language model has trillion-level token training, and the actual interaction data available for the embodied model is less than one-tenth of that.

Digital mining is about to explode, and market space will be fully opened

According to Future Markets data, the global physical AI market will grow from US$383 billion in 2026 to US$3.26 trillion in 2040, and will enter an explosive phase in the next few years. According to the 2026 Robotics Industry Link Conference, GPT-2 and GPT-3 training data correspond to about 790,000 hours and 11 million hours, respectively, and require at least 10 million hours of multi-modal data to achieve usable physical intelligence, while factors such as multiple scenarios, multiple modes, yield, long-tail data, and multiple participants have actually required far more than 10 million hours of multi-modal data sets, and Bee Finder has raised the target data collection volume to 10 billion hours in 2030.

Various digital acquisition routes such as complete machines, EGO, and UMI coexist. The marginal demand for high-quality data is becoming more and more determined, and cameras, posture perception, and tactile perception are becoming more and more important

Currently, the complete machine, UMI, and EGO (first person view) solutions are mainstream real data collection solutions. Among them, the whole machine data is difficult to share and the collection cost is the highest; EGO has the advantages of lightweight, low cost, and high versatility; UMI single equipment is expensive, but the data accuracy is high. No matter what kind of digital acquisition technology route is adopted, the future industry's demand for higher and higher data quality is determined, and there is more and more room for camera, attitude sensing, and tactile sensing applications.

The digital mining explosion is expected to drive a rapid increase in the scale of the following links

(1) Digital acquisition equipment: including complete robot machines, UMi and eGO equipment, attitude sensing IMU, dexterous hands, tactile sensors, cameras (2D+3D), VR glasses, etc. The digital acquisition of the whole machine requires 1:1 support for the robot itself and VR glasses; first-person data requires hand-eye coordination as the dominant requirement, requiring a camera headband and a dexterous hand for 1:1 support, respectively.

(2) Simulation platform: Simulation is currently one of the core training methods of physical AI companies, with obvious cost and data output advantages.

(3) Digital acquisition company: Digital acquisition company has strong experience in scenarios, data standardization and data labeling platforms. With data sales as its main business model, its core competitiveness lies in the cost and quality of data output.

(4) Vertical applications: Exclusive databases and digital acquisition scenarios in various industries are highly scarce. Vertical physical models are the most valuable racetrack for physical AI.

Risk Alerts

The development of robots falls short of expectations, and there is a risk of iteration in data collection technology routes.