By Sarah Xiao 
Start with the conclusion: Chinese industrial robotics has won volume. It has not won process.
In the first half of 2026 China produced 537,700 industrial robots, up 28 percent year on year; May was the first month above 100,000 units, June reached 110,700. Domestic brands held 54 percent of the Chinese market in 2025 and crossed 55 percent in the first quarter of 2026, with Estun taking roughly 10 percent and the top position across all brands. Robot density passed Germany and Japan, trailing only Korea and Singapore. Exports exceeded imports for the first time in 2025, and reached ¥6.29 billion in the first half of 2026, up 18.6 percent, into 141 countries.
All of that is real. All of it counts units, not process.
This piece is not about the slogan of substitution. It is about five specific exams: semiconductor cleanroom handling, body-in-white laser welding, force-controlled grinding, gigacasting tending, and offline programming with digital twins. Exam by exam: where the gap is, which part has closed, and how to split a purchase.
Exam one: semiconductor cleanroom handling
The entry ticket is four numbers: repeatability in the ±0.005 mm class, cleanliness down to ISO Class 1-3, particle and outgassing control, and mechanical plus communications docking with EFEMs and FOUP carriers.
The incumbents’ advantage is not the machine. It is that the whole line is built that way. When a fab is constructed, the handling units, interface standards, maintenance manuals and spare-part lead times arrive as one package. Replacing a supplier means replacing a validated process, not swapping a robot.
The domestic position: cleanroom models exist and are meaningfully cheaper; back-end packaging, panel and photovoltaic lines with looser cleanliness requirements already run them in volume. Front-end critical stations remain largely imported. That is not a capability problem so much as a validation-record problem — on a 300 mm line, every unplanned stop generates a report.
Exam two: body-in-white welding
BIW is where robot density peaks and where the big four still earn their steadiest margin. A mainstream welding line runs around 60 JPH with several hundred robots, mixing spot welding, adhesive dispensing, laser remote welding and inline metrology on one line.
The gap here is not repeatability. It is three things: matching fixtures and robots to the takt, closing the loop on weld-quality inspection, and how fast a fault is localised when the line stops. An hour of downtime is measured in tens of thousands of yuan.
The domestic opening is electrification: gigacast parts, chassis components, battery trays. Chinese robots are already in those lines at volume, where takt requirements sit one notch below a traditional main welding line and validation cycles are shorter. The main welding line stays imported for a plain reason — no OEM wants to carry validation risk on a new model launch.
Exam three: force-controlled grinding
This is the segment where domestic suppliers have closed fastest. A six-axis force sensor plus line-laser vision gives constant-force tracking and path compensation; aerospace structures, EV housings and casting deburring are the classic cases.
The hard part is not hardware. It is the parameter library: for each material, allowance and tool-wear state, how much normal force, what feed rate, how to back off when out of tolerance. Incumbents sell decades of accumulated parameters and commissioning experience. Domestic suppliers sell hardware plus an integrator who tunes it.
So the test is specific: not surface finish in a demo, but force drift and tool life over eight continuous hours. Domestic packages are sufficient for high-mix low-volume work. For high-volume, high-takt lines, ask for first-pass yield data.
Exam four: gigacasting tending
Electrification made gigacasting standard, and it created demand for heavy payload robots: 500 to 900 kg, high temperature, high takt, hard impact.
The bottleneck is not the arm. It is the RV reducer. Industry data puts domestic content at about 65 percent for strain-wave reducers and 72 percent for servo motors, but only 50 percent for RV reducers — the last hard bone in the component stack. Heavy joints are the most demanding application for RV stiffness, backlash and life. Until that closes, price and lead time on heavy models are set elsewhere.
Exam five: offline programming and digital twins
This is the exam most often skipped, and it sets the delivery cost of the other four.
The incumbents’ moat is a software ecosystem: tools that chain robot models, process packages, post-processors, cycle simulation and code generation into one path, with integrators accumulating reusable engineering assets on top. Domestic controller makers have all been building platforms, and progress is real, but the weak point is the standard-component library and post-processor coverage. Software is not drawn; it is fed by thousands of projects.
A practical test: hand over your part model and ask for a downloadable program with cycle simulation and collision check within 48 hours. Whoever can do that has offline capability.
Where the gap actually is
Stack the five exams and the gap concentrates in three places, none of which is “not accurate enough.”
Accuracy retention. Factory accuracy is calibrated; accuracy six months later is held by structure and compensation. Thermal distortion, wear, cable ageing, encoder drift — these appear over eight continuous hours. A large part of the incumbents’ premium buys “still the same number next year.”
Process software. Weld parameter libraries, adhesive path libraries, grinding force parameters, post-processors. These are not algorithms; they are engineering assets. Domestic suppliers started five years later at least.
Systems integration know-how. How to sequence takt, how to make fixtures cooperate, how to close the inspection loop, how to recover from a stop. Much of this sits with integrators, and integrator margins and payment terms decide whether they can afford to accumulate it.
How to split a purchase
| Scenario | Imported benchmark | Domestic status | Recommendation |
|---|---|---|---|
| Front-end semiconductor handling | ±0.005 mm class, full cleanroom certification | Cleanroom models available, thin front-end experience | Import for critical stations, domestic for back end and periphery |
| BIW main welding line | Takt and stability fully validated | Volume use on parts, chassis, battery trays | Import for main line, domestic for component lines |
| Force-controlled grinding | Mature parameter packages | Complete solution, thin parameter library | Domestic for trial production, measure first-pass yield before volume |
| Heavy gigacasting tending | Mature above 700 kg | Constrained by RV reducers | Check reducer supply before committing |
| Offline programming, digital twin | Complete software ecosystem | Platforms improving, weak standard-part libraries | Test with a 48-hour program request |
55 percent is unit share, not value share
Two more numbers belong next to the headline: the average price of a domestic six-axis machine is around ¥180,000, down 30 percent in five years, and utilisation at mid-to-low-end general-purpose capacity has fallen to 58 percent. Volume rising, price falling, low-end capacity idling — together those say the first phase of substitution, buying share with price, is close to finished.
The next phase is not about units. It is about the process package in each scenario: parameters, fixtures, inspection, software, service response. Build that and 55 percent can become value share. Fail, and 55 percent is just a larger red ocean.
For buyers, the useful question is not whether domestic robots work. It is which stations belong in which bucket: those where validation records decide survival (import), those where takt and cost decide survival (domestic), and those where both must be tested (domestic trial, then let first-pass yield decide).
For suppliers, the risk is not foreign price cuts. It is mistaking share leadership for capability leadership. The first is a financial statement. The second is the downtime number on a customer’s line.