When Unimate, the world's first industrial robot, went to work at a General Motors plant in 1961, its job was to move hot die-cast parts. It was the starting point of industrial automation: a machine repeating fixed motions in place of a human doing dangerous work. Robots have evolved quickly in the 65 years since. They spread across manufacturing floors for welding and assembly, and began sensing their surroundings with cameras and sensors. With artificial intelligence attached, they gained the ability to distinguish objects they had never seen and to plan a next move. Humanoids now run faster than people and turn somersaults.
Work that requires hands is another story. Screwing in a light bulb or fitting a part whose position shifts slightly each time means touching the object and then continuously adjusting the force and position of the fingertips. That is a different order of difficulty from repeating a set trajectory. In an experiment that fitted a multi-finger robotic hand to Apollo 2, a humanoid built by U.S.-based Apptronik, Google DeepMind's Gemini Robotics 2, unveiled in July, succeeded 92% of the time at unscrewing a light bulb. But screwing one in worked 36% of the time, tying a garbage bag 44% and sealing a zip-top bag 40%. DeepMind assessed precise multi-finger manipulation as still a hard problem.
Wang Xingxing, founder of Unitree and a leading figure in Chinese humanoids, pointed to the same limits at the World Robot Conference in Beijing last month. In a fixed environment, he said, repeated training can produce a high success rate, but performance drops sharply when the environment changes, and a new task requires training all over again. He put the industrial inflection point at machines that can carry out about 80% of tasks in an unfamiliar setting from voice or text instructions alone. Reaching that level, he said, would take two to three years at the earliest and five to 10 years at the outside.
On shipments, China is already far ahead. Global humanoid shipments topped 22,000 units in the first half of this year, according to research firm Counterpoint. The top five companies were all Chinese, and together they accounted for 86% of the total. Yet more than 60% of those units went to entertainment and performance or to data generation and research, while intelligent manufacturing took 13%. Shipments are a gauge of production capacity, not a measure of how much human labor is being replaced on industrial sites.
Korea has a wide range of demanding manufacturing sites for real-world testing, from shipyards and steel mills to semiconductor, battery and automobile plants. Robot density in Korean manufacturing stands at 1,220 units per 10,000 employees, the highest in the world, according to the International Federation of Robotics. Shipbuilding, which handles irregular structures; steel, with its high-temperature and heavy-load work; and advanced manufacturing processes that require precision measured in micrometers are all good places to test humanoid performance and reliability. They are sites where a motion that worked in the lab can be checked for whether it repeats at the same speed and accuracy on an actual production line, and where data can be accumulated.
More companies are digging into hands and precise manipulation. Samsung Electronics (005930.KS) this year set up Hand Lab, a team dedicated to robotic hands, inside its future robot initiative, and in July launched an RX business division reporting directly to the chief executive. Robot makers including Robotis and Tesollo are advancing precision gripping and manipulation technology.
In the 2027 budget plan approved at a Cabinet meeting on the 1st, the government allocated a record 39.5 trillion won to research and development. The physical AI budget rises to 3.1 trillion won from 900 billion won this year. Spending on that scale is unprecedented, and it should not end as a simple program to distribute robots. It has to push up the success rates and uptime of hand work such as precision assembly and tool use, and feed the force, tactile and behavioral data generated along the way back into technology development.
China's production capacity is formidable, but industrial competitiveness in humanoids is not settled by production scale alone. The more important standard is how reliably a machine performs the precise work human hands have handled, and how quickly it adapts to new environments and processes. In that area, no one is decisively ahead yet. Korea has demanding manufacturing sites for real-world testing, companies pursuing hand and manipulation technology, and now large-scale policy funding as well. There is no reason to burn those resources chasing China's output. Korea should convert the strengths of its manufacturing base into humanoid task performance and an edge in data. It is too early to call China's volume advantage a final victory.
