The IFA 2026 in Berlin featured the largest number of humanoid robots ever, with a notable presence from Chinese robotics companies. Firms such as Aizhi Robotics, Dubot, Engine AI, and DEEP Robotics showcased their innovations. Unitree's humanoid robot demonstrated its capabilities by assembling a sandwich using bread and vegetables.
In late January 2025, 16 Unitree H1 robots performed the traditional Chinese dance Yangge while distributing handkerchiefs during the popular CCTV Spring Festival Gala. In April of that year, humanoid robots participated in a 21.0975 km half marathon in Beijing, with the winning Tian Gong Ultra completing the race in 2 hours, 40 minutes, and 42 seconds. At that time, the robots required human guidance and operated in a semi-autonomous mode. This year, Tian Gong Ultra completed the marathon autonomously, reducing its time to 1 hour and 15 minutes.
The evolution from dancing robots to those capable of performing tasks marks a significant advancement in China's humanoid technology. However, it is premature to equate the robots showcased at exhibitions with fully operational 'working robots.' Many challenges remain before they can reliably function in industrial settings for extended periods.
Factories and warehouses do not operate under fixed conditions. Variations in component locations or task sequences can occur, and unexpected situations may arise, such as a person entering a robot's path. Simply repeating predetermined actions is insufficient to handle these variables. For robots to work reliably over long periods, they must be able to perceive their surroundings and respond appropriately, which requires accumulating diverse data from real-world tasks. This is part of the Chinese government's push to accelerate the deployment of humanoid robots in industrial environments.
In June, China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special initiative for 'humanoid robots and embodied intelligence practical training.' This initiative focuses on training and validating robots in real work environments. 'Embodied intelligence' refers to the ability of AI to perceive and act upon its environment through a physical body, closely related to physical AI.
The goal of this policy is clear: to generate data through practical experience. As robots pick up and move objects, they create various types of information, including video and location data, movement trajectories, force adjustments, joint movements, and task sequences. China aims to utilize this data to improve models, algorithms, and key components. By the end of this year, the government plans to identify over 100 high-value application scenarios and develop the capacity for 'mass-scale field application.'
Moreover, the deployment of robots is not limited to factories. Their applications extend to logistics, commercial services, healthcare, caregiving, safety production, and emergency response. Data accumulated in factories alone may not suffice for hospitals or disaster sites, as different environments require distinct judgments and actions. Expanding the range of applications not only broadens the market but also lays the groundwork for acquiring necessary data across various work environments.
However, it is crucial not to equate policy goals with current technological capabilities. 'Mass-scale' does not imply that such a number of robots have already been sold or deployed; rather, it signifies the ability to expand to that scale. The Chinese government has also identified challenges such as durability, thermal management, power consumption, collision detection, and emergency braking for long-duration, high-load tasks.
South Korea is also ramping up investments in this area. In August, the government announced plans to invest 2.3 trillion won in fostering a humanoid ecosystem by 2030 and an additional 2.8 trillion won for demonstrations in eight key sectors, including autonomous manufacturing, small-scale manufacturing, agriculture, construction, logistics, ports, caregiving, and defense. Furthermore, over 500 demonstrations are scheduled for 2027, with data from various sectors being consolidated into a government-wide integrated library for domestic companies' model training.
However, the most critical consideration for South Korea is not the scale of investment but the type of data to be secured in practical demonstrations. The country has already accumulated extensive production experience and process know-how in related fields such as automotive, shipbuilding, semiconductors, batteries, and electronics. If this data, including video, location information, force, torque, and joint movements, can be utilized for training rather than merely confirming the performance of completed robots, it could significantly enhance the value of its manufacturing base.
That said, merely increasing the volume of data can lead to limitations in its applicability. Each factory has different equipment and workflows, and companies use various robots and sensors. Even if data is generated from the same task, differences in format and units can hinder its use in other environments or models. While it is impossible to standardize all data into a single format, establishing minimal common standards for what information to record and in what units and structures is essential. It is important to create mechanisms that allow for the use of necessary information while protecting companies' technologies and trade secrets.
In the competition for generative AI, computational resources and digital data were crucial. In the next phase, where robots and AI converge, real-world work environments will be the key resource. This shift presents a new opportunity for South Korea, which has a strong manufacturing base. However, China is rapidly expanding robot development and deployment with robust policy support, making it challenging for South Korea to match that pace in the short term. Competing on scale in the same manner is also not realistic for South Korea. Therefore, the country must leverage its strengths in manufacturing to gradually accumulate advanced practical data and refine the process of connecting this data to technological advancements. It is time to build competitiveness not through scale but through the quality of the field and the depth of accumulation.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.