Rough Robots Learn First

Humanoid robots are not yet at the level where they can work like humans, but companies are beginning to deploy them in factories. As a result, the traditional sequence of finishing them first and then putting them into production is being reversed, toward deploying them on the production floor first. Rather than making them well and then using them, companies are starting to use them in order to make them well.

At IFA 2026 in Berlin, Germany, exhibitions featured humanoids dancing and performing somersaults, as well as autonomous robots playing soccer against one another. IFA described it as the largest humanoid exhibition ever. The main topic at IFA 2026 was AI-based "physical AI," which refers to perceiving and acting in the real world.

Beyond that stage, tests are also under way to verify real-world work performance. China’s humanoid robot "Tien Kung Ultra" recorded 8.64 seconds in the 100m, and its developers are testing it on carrying boxes in factories. In China, there have also been cases of robots for sorting mail and packaging smartphones being deployed in the field. The test stage is shifting from flashy movements to repeatable labor.

Georg Stieler, a robot analyst advising Chinese industry, estimated that 50%–70% of the humanoids produced in China this year will be deployed not on production floors but in "data factories" for collecting training data. This is because the Chinese securities industry believes that, if industrial humanoids are to recoup their investment versus labor costs within 2 years, including maintenance costs, they need to be at about CNY 160,000, while actual price analyses put industrial humanoids at typically CNY 300,000–500,000. This suggests that learning robots are proliferating ahead of robots intended to generate revenue.

That said, this should not be seen simply as a limitation of an immature industry. Physical AI requires data obtained by moving a body in the real world, and learning data accumulates from experiences such as missing an object while trying to pick it up, encountering unexpected obstacles, and repeating the same motion. These records are fed back in as training material for AI, and the process of making and operating robots itself leads into the development of the next robots.

China’s competitiveness in humanoid robots should be read in a different sense from product perfection. First comes the possibility of mass production, based on a dense supply chain surrounding the motor, reducer, actuator, sensor, battery, and semiconductor components that make up humanoids. As robot production expands, opportunities for real-world deployment increase, and as real-world deployment expands, opportunities to accumulate trial and error also rise. Accordingly, the core of the China threat is interpreted not as humanoid manufacturing capability itself, but as the conditions for making many, moving many, and failing many times.

According to a USCC analysis, China has a structure that systematically secures corporate and industrial-site operating data, and it also has a structure that systematically secures data from the physical world. USCC analyzed that this structure could become an advantage in autonomous driving AI development. It also said the same structure could become an advantage in humanoid AI development. The flow is that supply chains and data-acquisition structures can support real-world deployment and learning accumulation, leading to strengths in AI development.

Hyundai Motor Group has chosen the factory as its testing ground. Hyundai Motor Group plans to deploy Boston Dynamics’ humanoid robot "Atlas" at Hyundai Motor Group Metaplant America (HMGMA) in Georgia from 2028. "Atlas"’s first task will be sorting parts by process sequence. Hyundai Motor Group also plans to build a robot production facility in the United States with an annual capacity of 30,000 units. Through this, it aims to connect a single value chain spanning development, learning, verification, mass production, and service operations.

Factories are a favorable environment for robot learning. That is because work can be controlled, movement paths can be controlled, and the same task can be repeated. In that sense, what stands out in the combination of Hyundai Motor Group’s manufacturing sites and Boston Dynamics is not the number of robots deployed inside the factory itself.

The next variable in competition is how tightly robot manufacturing capability is linked to robot work experience. This raises the same challenge for Korean manufacturing. Korean manufacturing has semiconductor production capability, battery production capability, motor production capability, sensor production capability, and final assembly capability. But those capabilities alone do not guarantee an edge in physical AI. The necessary condition is converting hardware production capability into the ability to accumulate real-world data, and converting hardware production capability into the ability to improve AI.

In the end, the nature of the humanoid race is not a race to launch finished products first. It is closer to a race to deploy them in the real world first and turn failures into learning. Even while waiting for fully finished robots, immature robots are already learning to work.

Source: IT DAILY · Kim Byeong-ju
Original: https://www.itdaily.kr/news/articleView.html?idxno=241645

References

• https://www.itdaily.kr/news/articleView.html?idxno=241645