At 9:20 p.m., Zhao Xin logs onto Douyin for another livestream. On the previous broadcast, a viewer said a robot dog’s knee joint was making an unusual noise. Zhao planned to open up the knee motor of one leg on camera, showing viewers how to disassemble and reassemble it while also giving that owner a remote diagnosis.
He says certain sounds now point to familiar problems. A rustling noise usually means the unit needs lubrication. A clicking sound often points to damaged gears in the planetary reducer. A buzzing electrical sound may signal a short circuit on the circuit board.
His repair-focused livestreams draw more than 1,700 viewers per session, with more than 40 watching at the same time. The audience is not large, but many stay engaged, and some regulars return every day.
The appeal is tied to a broader wave of public attention around robots. Videos of robots failing in public have spread widely online, including machines carried away after a marathon, one losing its head in a cage fight, and others collapsing during performances.
More robots are creating demand for maintenance and repair
The installed base is rising quickly. IDC data cited in the report put global humanoid robot shipments at about 18,000 units in 2025. A Ministry of Industry and Information Technology official projected that China’s output of complete humanoid robots could exceed 100,000 units in 2026.
As more machines are deployed, demand for maintenance and repair is also increasing, and early movers have already entered the market.
A self-taught repair worker built a following by taking machines apart live
Zhao works at a robot and drone repair training institution in Ordos, Inner Mongolia. On Douyin, he presents himself as an ordinary person learning robot repair from scratch. That framing is close to his actual background.
Born in 1990, Zhao studied civil engineering at junior college level and spent most of his career in service jobs. He liked taking apart electronics, but had never worked in repair before. Most of what he knows about fixing robots came through self-study.
He first opened up a robot dog only about half a year ago. The first time he disassembled and reassembled one, he ended up with 10 extra screws and had no idea where they belonged. He had to open a second robot dog and compare the parts one by one.
The first time he worked on a humanoid robot, he snapped three screws and broke two clamps. Some screw holes had been reinforced with anaerobic adhesive. After using a heat gun to soften the glue, the screws themselves also softened, and they broke under force.
In another case, a robot he was repairing tried to stand up from a squatting position. It should have pushed forward, but instead pushed backward and fell into a split. After checking the machine, Zhao and his colleagues found that the square inertial measurement unit, or IMU, chip in the robot’s waist had been installed 180 degrees the wrong way around. The software showed no fault, but the force direction was completely reversed.
Routine maintenance matters as much as repair
Robot dogs and humanoids need upkeep, not just repairs. Zhao described one robot dog that showed no error in the app but could not stand steadily. After taking it apart, he found that a bearing inside the leg joint linkage had worn down and developed two deep grooves because maintenance had been delayed.
Even after those early mistakes, Zhao keeps repeating the same point: the technical barrier to robot repair is not especially high. The first robot dog he fixed had a communications fault. Replacing a cable and doing some soldering solved the problem in an hour.
In many cases, a technician can use the customer’s description and the error messages in the app to identify the likely location of a failure. Once the faulty component is found, the job often comes down to replacing the part rather than repairing it in place.
Practitioners interviewed for the report broadly said humanoid and robot-dog repair can be easier than fixing robotic vacuum cleaners or drones.
The parts mix is different as well. Compared with robotic vacuum cleaners, humanoid robots and robot dogs use relatively straightforward assemblies, mainly joint modules, perception systems, mainboards, and batteries. The joint modules include motors, reducers, encoders, driver boards, and bearings. Perception systems include radar, cameras, and IMUs.
The hands-on techniques are also less delicate than some drone jobs. The report points to a repair method known as “flying wire,” where a 0.2 mm copper wire is looped at a solder point to restore the connection between a chip and the circuit board. That process is done under a microscope and demands careful manual control and patience.
Zhao said someone starting from zero can become job-ready in a little over a month. People with electronics and circuit knowledge or prior repair experience tend to learn faster.
Parts access, not skill, is the main bottleneck
For Zhao, the hardest part of the business is not the repair work itself but the control manufacturers have over spare parts.
Core robot components are typically customized by the manufacturer, not interchangeable across brands, and there are almost no third-party substitute parts. Authorized distributors can buy components from manufacturers, but those purchases are often limited by region and intended use. Shops without authorization have to look for used parts from dismantled machines or develop their own channels.
Repair operators also lack the documentation available in other electronics categories. When fixing phones or drones, technicians can often refer to public circuit diagrams to identify chip functions and line connections. For robots and robot dogs, those diagrams are generally unavailable. In some cases, the report says, chip names or markings are intentionally removed before the product leaves the factory.
Peripheral items are costly too. Zhao said replacing the rubber soles on a robot cost 400 yuan for one pair, while his own shoes cost 59 yuan.
Last year, the institution where he works bought a suit for a robot to wear in a performance, and that alone cost 800 yuan. According to the report, some garment factories have already started producing clothing and accessories designed specifically for robots.
Third-party repair shops are targeting out-of-warranty machines
During one livestream, Zhao spent more than 10 minutes removing one leg from a robot dog. He laid the parts out neatly on a table and explained their structure, function, and operating principles one by one.
The model was Unitree’s Go2. The Go2 series and Unitree’s G1 humanoid are the models third-party repair shops and training institutions see most often.
In 2025, Unitree led humanoid robot shipments with more than 5,500 units sold, while AgiBot ranked second with more than 5,100 units, according to the report. The gap between those two and the rest of the field had already widened.
Shipment growth is still accelerating. A Unitree executive predicted the company’s total shipments in 2026 would at least double from the previous year and could even rise by nearly three times. AgiBot said in late June 2026 that cumulative output of its general embodied robots had passed 15,000 units.
Unitree also leads in robot dogs. Founder Wang Xingxing said in July 2025 that Unitree’s quadruped robot sales accounted for 60% to 70% of global shipments.
Sales of robot dogs are higher than those of humanoids. Data from GGII, the Gaogong Robot Industry Research Institute, showed that China sold about 57,600 quadruped robots in 2025, nearly three times the humanoid total.
As the installed base grows, brands beyond Unitree are also generating work for independent repair shops. Fang Jinghua, general manager of Nanjing Kaogong Yunji Robot Technology, said his company operates five repair points nationwide. Since the start of this year, orders involving robots from AgiBot, Leju, Zhongqing, and other brands have gradually increased.
Commercial deployment remains at an early stage, but usage patterns are starting to diverge by brand.
Unitree and AgiBot, the two volume leaders, both market general-purpose platforms and all-scenario applications. In practice, Unitree’s humanoids are used more often in research and performances, while its robot dogs are common in inspection, surveying, and emergency response. AgiBot’s main landing points are industrial manufacturing, commercial retail, and cleaning inspection.
Among brands with lower sales, Leju is focused on data collection and research. UBTech is oriented toward industrial scenarios. Booster Robotics is used mainly in competitions and research. Songyan Power’s smaller robots are used primarily in entertainment performances. Fourier’s most distinctive use case is healthcare and elderly care. Galaxy General focuses on commercial retail, industrial manufacturing, and cultural tourism.
The report notes that each brand’s official positioning spans multiple scenarios, and the article lists only the more representative or distinctive real-world uses.
Mainstream robot and robot-dog models usually come with a warranty period of about one year. During that period, most users choose to return them to the manufacturer for repair. Manufacturers generally state that non-human-caused quality issues are repaired free of charge during warranty, while damage caused by users, private disassembly, or failure to follow instructions is billed to the customer.
Independent repair becomes more competitive once machines are out of warranty. Fang said his company’s average repair bill is typically 10% to 15% of the robot’s selling price and can come in 20% cheaper than factory service. A Shenzhen merchant cited in the report said some jobs can be done for half the cost of return-to-factory repair.
Timing is another selling point. Factory repair cycles generally run longer than one month, while third-party providers can often finish in about a week.
Research is the largest use case, but performance damage drives many repair orders
According to the 2025–2026 Humanoid Robot Industry Development Research Report released by the Humanoid Robot Scenario Application Alliance, education and research account for 42% of application scenarios. Data collection, services, and performances follow, while industrial use has the smallest share.
But the robots arriving at third-party repair shops are often damaged during entertainment performances. Customers are usually robot rental companies, self-media creators, and other small independent operators.
The business also has a seasonal pattern. The period before the May Day holiday is a peak season because new merchants open and entertainment demand rises. Summer is slower because outdoor heat reduces commercial events.
Institutions using robots for research, development, or data collection are usually less sensitive to price and more willing to return units to the manufacturer.
For now, repair alone does not pay the bills
Even with lower prices and faster turnaround, third-party repair remains a hard business to monetize on its own.
The biggest reason is simple: there are still too few serviceable robots in circulation. Many units remain under warranty. Among those that are out of warranty, a substantial share still goes back to the manufacturer. Add in geographic dispersion, and the number that reaches local third-party shops stays limited.
Zhao’s institution receives only one to three repair orders a month for robots and robot dogs on average. Fang’s company, which operates five repair locations nationwide, repaired more than 10,000 household robot vacuum cleaners and various commercial service robots over the past year, yet it handled only five or six humanoid repairs and a little more than 10 robot-dog repairs per month on average.
That is why shops in this segment all have other core businesses. Zhao’s employer, Zhongyun Zhike (Inner Mongolia) Technology, was built by an owner who started with machinery operation training related to open-pit coal mines. Fang’s company makes money from repairing robotic vacuums and commercial service robots. Shenzhen Leike Intelligent Technology has years of experience in industrial robot repair. Beijing Silicon Park Technology relies on robot rentals, study tours, and related business. Some other institutions focus mainly on drone repair and training.
Training has moved in early, with entry-level pay around 6,000 to 8,000 yuan a month
To add revenue, several of these organizations have launched, or are planning to launch, training programs for robot and robot-dog repair.
The training market is still early and lacks a common standard. Providers vary sharply in background, course length, and tuition. Most teach similar content, including fault diagnosis, disassembly and assembly, component replacement, and calibration. Programs range from eight days to 40 days, and tuition ranges from 5,000 yuan to 30,000 yuan.
Lu Anding, head of Shenzhen Leike, said six training cohorts run since February this year have produced more than 100 trainees. About 70% to 80% of them entered the robotics industry after graduation.
Most of those graduates did not go into repair specifically. They took entry-level jobs in sales, operation, and maintenance.
Lu used robot rental companies as an example. Someone has to transport the machines to event sites, operate them, and monitor the full performance. By his conservative estimate, every three robots require one operator. Manufacturers cannot train that many people on their own, so outside institutions are stepping in and then referring workers to employers.
Interviewees broadly said the talent threshold for repair work is not high. Graduates from junior colleges and vocational schools can usually do the job, and some training institutions have lowered the recruitment threshold to high school education.
That changes when the work expands from basic repair into secondary development. Those roles usually require a bachelor’s degree because they involve more specialized robotics knowledge as well as programming and algorithms.
Interviewees gave a broadly consistent pay range for the repair segment: newcomers usually earn about 6,000 to 8,000 yuan a month, while experienced workers can make more than 10,000 yuan, with regional wage levels affecting the exact figure.
The report also says large companies such as JD.com have already started laying groundwork. JD says it plans to train 100,000 engineers within five years, covering robot repair and related services. A number of vocational schools are also preparing to open robot repair majors.
Still, regardless of who provides the training, robot repair remains a business that has to wait for the broader market to mature.
The article was originally published by the WeChat account QbitAI and written by Lin Fangzhou.