Separating two pieces of fabric may look simple. For a humanoid robot, however, the soft, slippery material is a demanding test of dexterity: too much force can distort it, while too little leaves it beyond the robot's grip.

Yet in Chinese garment factories, robots developed by Aitu, an intelligent sewing technology company based in east China's Zhejiang Province, are already being tested for separating fabric pieces, lifting them, laying them flat, and aligning them for sewing.

For a single product category and fabric type, the company says its robots have achieved a 97% success rate in separating fabric pieces, with more than 98% of the resulting sewing work meeting quality standards.

At the current level of efficiency, Aitu estimates the robot could pay for itself in about 18 months.

The calculation goes to the heart of the industry's next test: whether humanoid robots moving from laboratory prototypes to factory floors can deliver productivity gains large enough to justify the investment.

The industry widely regards 2026 as the year humanoid robots begin moving into mass production. China's Ministry of Industry and Information Technology and its state assets regulator have launched a program to train humanoid robots and embodied AI systems in real-world settings, aiming to move them from testing into actual operation and build deployment capacity at the 10,000-unit scale by the end of the year.

Aitu robots are being deployed and put through production trials at Chinese garment manufacturers including Sunrise Group, Eifini and Chenfeng Group. Their assignments remain narrow: loading and unloading fabric at template sewing machines, positioning material for pocket setters, and picking up and placing cut pieces at fusing press machines.

"Humanoid robots are still at the learning stage," said Lou Lingtong, general manager of Aitu. "They can perform individual processes well, but are still unable to run an entire production flow without human involvement."

Another industrial embodied AI robot, the G2 developed by PIABOT Robotics, has been deployed in electronics assembly, logistics sorting, component inspection and production-line loading.

In one trial at a major Chinese electronics manufacturer, the company says the robot completed 2,283 tasks over eight hours without an error.

On an automotive safety-equipment line, a G2 robot was put to work tapping holes in aluminum seat-belt components. The task had previously taken a worker about 18 seconds per cycle. The robot cut that to as little as 12.9 seconds, improving efficiency by roughly one-third.

These machines are not yet general-purpose workers capable of handling an entire production process. They are entering factories one station and one repetitive task at a time. Yet for manufacturers struggling to recruit and retain workers, even limited automation can be commercially meaningful if it is stable, repeatable and economically viable.

Lou said the number of garment workers retiring or leaving the industry each year is far greater than the number of jobs currently being taken over by robots.

"Factories can't recruit enough people or persuade them to stay," Lou said. While robots may initially fill labor gaps, Xiong Rong, chief scientist at the Zhejiang Humanoid Robot Innovation Center and a professor at Zhejiang University, said their wider use would also create jobs in maintenance, debugging and secondary development.

The next challenge is giving the machines enough real-world data to learn, adapt and transfer their skills reliably to factory floors.

In Cixi, a manufacturing hub in Zhejiang, a humanoid robot training facility was launched in mid-May with 40 embodied AI robots deployed. Designed around simulated real-world scenarios from nine local industries, including home appliances, automobiles and textiles, it is expected to provide an average of more than 90 hours of effective robot training a day once fully operational.

Similar efforts are taking shape in Hangzhou, where a national pilot base is linking robot hardware, AI models and industrial applications.

Such facilities reflect a growing recognition that access to real-world factory data is becoming critical to the industry.

Walking, visual navigation and route planning are becoming more mature. The harder problem is manipulation: grasping, turning, sorting and assembling objects repeatedly without sacrificing precision. For garment robots, flexible materials remain especially difficult.

"Handling fabric is like picking up tofu," Lou said. "Use too much force and it deforms, while use too little and you can't hold it."

The broader challenge is generalization: a robot trained to handle one material may struggle with another it has never encountered. Lou said enabling robots to adapt their dexterous movements to unfamiliar materials and objects is a key technical problem the industry is working to solve.

Industrial production leaves little room for error. Reflections and dust can disrupt perception, while continuous operation places heavy demands on durability, heat management and component life. A lack of common interfaces and safety standards further raises integration costs.

For now, the business case remains confined to narrowly defined jobs. Lou expects 2026 to 2028 to be a period of validating robots in individual processes, before coordinated deployment across multiple tasks begins to scale between 2028 and 2030.

Mass adoption is unlikely to come in one dramatic leap. It is expected to advance one workstation at a time, as robots prove they can operate through a full shift with the speed, precision and reliability needed to justify their cost.