China Galaxy Securities: Increased supply of digital acquisition solutions to compete for upgrades in commercialized applications

The Zhitong Finance App learned that China Galaxy Securities released a research report saying that the embedded robot industry is still in the 0 to 1 stage. The future industrial level is progressing positively, and it is recommended to continue to pay attention to its investment opportunities. The T-chain remains true without falsification, and the “real” supply chain still has investment value even after adjustment. The domestic chain gives priority to high-quality assets to be listed, followed by core component suppliers with a good structure and high value. It is recommended to actively pay attention to the main body manufacturer, and secondly, it is recommended to pay attention to parts with a good pattern, such as machine vision and harmonic reducers.

The main views of China Galaxy Securities are as follows:

Incidents

On July 17-20, 2026, the World Artificial Intelligence Conference (WAIC) was held in Shanghai. Many manufacturers of smart devices, models, and components participated in the exhibition. The bank believes that how to convince the market of commercial implementation capabilities is an important issue for current enterprises. This year's live demos tend to be more detailed, more generic, and long-range scenarios; at the same time, the mainstream model trend evolved from last year's VLA to “world model+VLA”, and the demand for EGO/UMI digital acquisition equipment and the number of suppliers have increased markedly. The following are specific observations.

Main body side

The number of companies and products has increased. By the first half of '26, the number of complete machine products in China reached 400, 70 more than at the end of '25. On the one hand, the current prototype production is not difficult; it is difficult to ensure consistent large-scale mass production and improve reuse life on the hardware side, as well as large-scale models on the software side. On the other hand, since the technical path has not yet converged and the standards have not yet been unified, the model factory and the main body manufacturer have penetrated each other. The model factory has launched its own body products as a carrier for presenting model effects, and the main body factory has also developed its own sizable models.

Model side

Most of the models shown last year were VLAs. Most of the world models shown this year were world models+VLAs, but each manufacturer's positioning of the world model was different. Some were built into the model to learn the laws of physics and preview before making decisions, and some developed the world model and VLA together as independent models.

Data side

1) Due to changes in digital acquisition methods, there are more diverse types of digital acquisition equipment. Last year, there were more motion capture devices. This year, there were more tactile gloves, UMI grippers, and Ego devices (such as 2 grippers, 1 headset, 3 mobile phones). Some manufacturers have ordered thousands of units for actual devices with the EGO and UMI solutions, and prices are showing a trend of falling pressure. It is not difficult to make a set of UMi/eGO digital acquisition equipment, but how to improve data efficiency and data modal dimensions is difficult. 2) In terms of training data structures, most of them currently use a 5-layer data structure of Internet data+human behavior data (human motion capture and ego data) +simulation data+real machine remote operation data+real world robot autonomous operation data, but there are differences in ratio. Currently, the importance of real machine data is still indispensable in post-training.

Application side

The highlights are refinement, higher generalization requirements, and the increase in long-term scenarios, and logistics sorting scenarios are expected to be the first to be implemented in real life. 1) There are more applications on the industrial side. Last year, they were basically handling, loading and unloading, and a few logistics sorting. This year, there were more sophisticated operation scenarios, such as Galaxy general robot screwing, flexible handling & precise assembly of Mecamand wiring plugs, and Lingchu Intelligence and Changfei Optical Cooperate with Changfei Optical Fiber to plug, inspect, vacuum package and pack optical modules. 2) On the commercial service side, many companies have shown scenes such as shelf retrieval (such as pharmaceuticals, retail; companies such as Galaxy GM, Mercamander, and Ant Afu), hotel laundry (companies such as Daxiao Robotics, Qinglang Intelligence, Qiongche Intelligence, etc.), and drink production. Shelf retrieval is generalized, and hotel laundry shows long-range task capabilities. 3) On the home side, the Galaxy General Robot showed the ability to cook breakfast, such as baking bread, and Fourier showed human-computer interaction in the home scene with a large external language model.

Risk warning: Risk of key technology breakthroughs falling short of expectations, risk of downstream scenario development progress, demand space falling short of expectations, etc.