The manufacturing industry seizes the opportunity for the large-scale commercialization of embodied intelligence.

"Embodied intelligence has different needs and expectations in the C - end (consumers) and B - end (enterprises). It's not easy to simply say which scenario has greater potential. If I have to choose, I think the B - end scenario will be implemented more quickly than the C - end."

When talking about the commercialization potential of embodied intelligence, He Chuan, the R & D director of Junpu Intelligence Group, told Jiemian News that from his experience with several leading robot manufacturers in recent years, they are all looking for a general way to solve the problem of reusable embodied intelligence. "Currently, it's difficult to find reusable scenarios in both C - end and B - end scenarios. The flexible modular process in the manufacturing industry may be a potential opportunity."

In the past year, embodied intelligence has become a new trend. The general public has witnessed the lively scenes of robots dancing, fighting, and running marathons, while practitioners are more concerned about the commercialization of embodied intelligence.

In March last year, Zhu Xiaohu, the managing partner of GSR Ventures, poured cold water on the embodied intelligence industry during a media interview. He said, "Who would spend more than a hundred thousand yuan to buy a robot to do these jobs?"

According to Jiemian News, the applications of embodied intelligent robots in scenarios such as greeting guests, reception, exhibition hall explanations, and shopping guidance mostly remain at the conceptual stage. However, in the industrial manufacturing field, some enterprises have begun to calculate the ROI of applying embodied intelligent robots to production lines, hoping to seize the opportunity for large - scale commercial use of embodied intelligence.

How can embodied intelligence move from the laboratory to the production line?

In late December last year, the industrial intelligent robot G2 jointly developed by the Junpu Intelligence - Zhiyuan Joint Laboratory started "on - the - job" testing on the production line.

The test results show that in micron - level contact - type flexible assembly, the success rate of this robot is about 99%, the success rate of reinforcement learning placement is 99.33%, and the success rate of visual grasping is 100%. Currently, the average operation cycle is 15.28 seconds, and the fastest cycle can reach 12.97 seconds.

An insider from a leading embodied intelligent robot company told Jiemian News that technological breakthroughs are the key for embodied intelligent robots to move from the laboratory to factory production lines, and flexible production capacity is a crucial indicator.

Flexible production capacity is the core ability for the manufacturing industry to transform from large - scale standardized production to multi - variety, small - batch, and personalized production. It requires the production line to be able to switch quickly and at low cost while maintaining stable output. Previously, flexible production capacity was widely used in the clothing, fast - moving consumer goods, and food processing industries. However, in recent years, industries such as automobile manufacturing, 3C electronics, and medical devices have also had an increasing demand for flexible production capacity.

As a provider of intelligent manufacturing equipment solutions and industrial digital software services, Junpu Intelligence mainly serves three industries: intelligent electric vehicles, healthcare, and high - end consumer goods.

He Chuan also noticed the changes in the production line requirements of these industries: In the traditional automobile industry, the vehicle platform is usually updated every 7 - 10 years, and the products produced during this period don't undergo significant adjustments. However, the vehicle platform of new energy vehicles needs to be completely changed every 1.5 - 3 years. The entire industrial chain, from upstream automakers to Tier 1 and Tier 2 suppliers, is facing rising hardware and software costs. The flexible modular process can solve these pain points.

He Chuan said that the flexible module is a process module nested on the basis of flexible production capacity. "There are mainly 72 types of assembly - related processes. Even if the product types need to be changed quickly, as long as the assembly process remains the same, we can reuse each module like building blocks, rather than simply disassembling the hardware and building a new production line. This will greatly improve efficiency."

More intelligent manufacturing enterprises have recognized the demand for production line flexibility in new energy vehicle manufacturing. Among them, Tianqi Automation, which focuses on automobile automation equipment business, is developing solutions for embodied intelligent robots in automobile manufacturing and other industrial manufacturing scenarios, and has invested in building an industrial data collection and training center project for embodied intelligent robots in Wuxi; Zhucheng Technology, which specializes in the R & D and production of automotive electronic connectors, is cooperating with Ubtech to explore the application of intelligent service robot parts wiring harnesses and connectors.

In He Chuan's view, embodied intelligent robots are expected to solve the industry pain points of traditional automation, such as insufficient flexibility, low reuse rate, and weak generalization ability.

Embodied intelligent robots are difficult to replace humans in the short term

In the past year, domestic embodied intelligent robot companies such as Zhiyuan, Unitree, Ubtech, and Galaxy Universal have been intensively competing in multiple aspects, including financing, shipment volume, and technical capabilities. Each company is trying to prove its competitiveness in the market.

However, the insider from the leading embodied intelligent robot company told Jiemian News that the embodied intelligent robot industry can't really be considered in a state of competition because no mainstream technological school has been formed in the directions of hardware, software, and models among manufacturers. "Most of them are still exploring in the laboratory scenario rather than in the real mass - production stage."

Currently, 200 units of the G2 robot launched by the Junpu Intelligence - Zhiyuan Joint Laboratory have been off the production line for testing. He Chuan revealed that although the accuracy of the G2 robot in several tests has reached over 99%, the generalization ability of the model still needs to be improved. After the product being produced is changed, even if the process is the same, the robot still needs a lot of time for model scheme verification, data annotation, learning, and training.

"Currently, embodied intelligent robots can't really replace humans. They need to enter the production line formally and operate stably for a period of time before evaluation. At present, the first stage hasn't been completed." He Chuan said that the G2 robot hasn't fully achieved stability in grasping raw materials. It's easy to pick up adjacent parts when grasping the target material. His team plans to continue the second - stage R & D in 2026.

According to Jiemian News' multiple sources, although the embodied intelligence industry has reached the stage of attempting industrial commercialization, the model capabilities are still far from being as mature as those of large - language models.

On the one hand, the models of each robot based on the "big - and - small - brain hierarchical architecture" are developed under the structural constraints of specific bodies. Model training and optimization highly depend on the hardware design, software architecture, and sensor configuration of the body, resulting in insufficient universality of the model body.

On the other hand, since the models are limited by different bodies, in most cases, real - machine data collection must be carried out for specific bodies, and it's impossible to use the general data sets obtained in the public domain like large - language models.

Lin Lin, the vice - president of Lenovo Group and a partner of Lenovo Capital and Incubation Group, once said when talking about this topic with Jiemian News and other media that with the popularity of embodied intelligence, people's expectations for its commercial application have increased. However, he believes that it will take 10 years to find the final answer to this problem. During this process, different application scenarios and different business development stages will emerge. The current leaders don't necessarily mean they will be the final winners.

In He Chuan's view, the period from 2026 to 2028 will be a crucial window period for the development of the embodied intelligence industry. "Currently, robots seem clumsy, mainly limited by computing power, data, and model capabilities. Among them, short - term, low - cost, and high - efficiency data simulation, synthesis, and collection will be one of the core topics in the next three years."

This article is from "Jiemian News", author: Xiao Fang, editor: Wen Shuqi. Republished by 36Kr with permission.