Within 1 to 2 years, robots will autonomously complete simple tasks among dangerous operations. In 3 to 5 years, robots will perform complex hazardous tasks on their own, with humans supervising only in exceptional situations.
Yan Weixin, chief scientist at the Shanghai Jiao Tong University Institute of Artificial Intelligence (AI), said in an interview with ChosunBiz on the 4th that this is when Humanoid Robot can replace humans at dangerous manufacturing sites.
Yan, the chief scientist, designed Shanghai's AI development plan and served as a Commissioner at a key laboratory of the Ministry of Industry and Information Technology (MIIT), which oversees China's high-tech industrial policy. The Shanghai Jiao Tong University AI Institute, where Yan currently works, plans Shanghai's AI development blueprint and carries out a state-led special-purpose robot project, serving as a core forward base for China's AI policy.
AgiBot, the robotics company Yan co-founded, shipped 5,168 Humanoid Robot units last year, ranking No. 1 globally.
As recently as a little over a decade ago, China remained on the fringes of industrial robots, but it now leads trends in the global Humanoid Robot market. Moving beyond past methods that relied on massive subsidies, it achieved overwhelming cost competitiveness by densely concentrating the value chain of core components in vast industrial parks.
On top of that, China is evolving rapidly thanks to a distinctive culture of innovation that boldly trains robots with wild data from real industrial sites—where variables pour in—rather than in controlled greenhouse laboratories.
There are currently as many as 160 Humanoid manufacturers in China, and nearly 10,000 corporations engaged in robot-related businesses, including 600 core component suppliers. The following is a Q&A with Yan, the chief scientist.
—What is the secret to the explosive growth of China's robot ecosystem?
Powerful collaborative synergy across the entire value chain, a dense supply network, and a strong will to invest in basic research. In massive high-tech manufacturing clusters such as the Yangtze River Delta centered on Shanghai or the Pearl River Delta around Shenzhen, you can find the entire supply chain—mold making, injection molding, precision machining—within the region.
This concentrated infrastructure has led to groundbreaking cost reductions. Another key is that we have trained algorithms with vast, living data acquired from chaotic real-world manufacturing and daily environments.
—What are the weaknesses?
We have achieved localization and stable supply at roughly 90% for structures such as robot bodies and 80% for batteries, but dependence on imports for high-end core components remains high.
The localization rate of torque sensors, which implement delicate sensing, is under 10%, and the planetary roller screw, a core drive component for knee and hip joints, also has low Production yield, putting localization at around 20%. The low Production yield and localization rate of these components are acting as bottlenecks to cost reduction.
—What tasks related to robot development are you currently focused on at Shanghai Jiao Tong University?
First is the manipulation capability of a versatile robot hand (dexterous hand) that combines vision and touch. We must implement actuators with 27 degrees of freedom in a space the size of a human hand.
Tactile sensing in particular is extremely difficult. Just as humans can grasp and handle objects with eyes closed, robots must instantly perceive pressure, texture, and slip, fuse that with visual information, and perform complex manipulations.
Second is precise control capability; third is implementing muscle memory and reflexes. Instead of calculating joint angles every time, we are focusing on studying reflex mechanisms that, by combining the brain's feedback, achieve complex movements and balance with minimal energy and computing power.
—Do Humanoid Robot need to closely mimic the human body, such as having five-fingered hands?
Looking at the course of human evolution, the lower body's walking ability may lag behind quadrupeds, but the upper body, which coordinates with the brain—especially the versatile "hand"—excelled in manipulation, ultimately making us the lords of creation.
Every object in our world—bottle caps, doorknobs, tools—has been thoroughly designed to fit the "human hand." The five-finger structure is the most ideal multipurpose data collector in environments that, like everyday life, are not standardized. It is also essential to implement grips that grasp or handle unexpected objects, enabling responses to diverse variables.
—Some say the devices are expensive and still fall short of human work capabilities, making it premature to adopt them in real industrial settings.
Whether robots are suitable for industrial sites is a calculation strictly based on economically rational behavior. If a robot is meant to replace the harsh high-temperature environment of a steel mill that reaches several hundred degrees, adoption is justified even if the device is expensive.
Conversely, for a robot that only tightens screws in a routine factory, there is no need to use top-tier components capable of flipping backflips. A uniform benchmark of what percentage of human capability must be achieved to be cost-effective is also meaningless.
Threading a needle is very difficult for robots, but in logistics, picking and carrying items can be done by robots far more precisely and efficiently than by people. We must look at value-to-price in the specific application scenario, not absolute price figures or uniform performance assessments.
—How is the process of Humanoid Robot replacing humans in hazardous industrial sites progressing?
The deployment roadmap is divided into three stages. First, autonomy begins with simple tasks among dangerous operations. Within 2 to 3 years, a model will be introduced where robots handle tasks and humans intervene via remote operation when emergencies or special responses are needed.
In 3 to 5 years, robots will make their own decisions and perform even complex tasks, evolving into a structure where humans intervene only in extremely exceptional situations with few data samples.
—Korea is considered favorable for robot training because manufacturing, including shipbuilding, is advanced.
Shipyards are a very promising stage where robots can learn and excel. The core process of welding involves such harsh working conditions that steel plates get unimaginably hot in summer. Moreover, unlike automobiles, ship components are less standardized, and work must pass through narrow, uneven spaces.
Such irregular and harsh environments are highly suitable for training "embodied AI robots" (embodied AI—robots equipped with AI that sense, learn, and adapt to the environment like humans). This is why on-site data from Korean shipyards are so important.
China, too, is currently making active attempts to introduce embodied AI robots for welding, grinding, and polishing in shipbuilding and repair processes.
—Where do you think Korea leads in the field of Humanoid Robot?
Korea has a major strength in reducers, the core of robot joints, especially in harmonic reducers. It has achieved pioneering innovations for over a decade, and the kinematic design capabilities currently demonstrated by several universities remain world-class and excellent.