This article is based on publicly available materials, intended for information exchange only, and does not constitute any investment advice
Last week, I visited Jining, Shandong, to tour the manufacturing facility of a listed robotics company.
The company is Rokae Robotics, listed on the HKEX. It excels in core control technology for robotic arms and is an unavoidable name in China's industrial robot sector. B2B business traditionally prioritizes profit over fame; combined with its listing on the HKEX, its public recognition naturally lags behind names like Unitree and AgiBot (Shangwei New Materials), which frequently make headlines.
The organizer of this event was JD.com. The e-commerce giant we know well may also become a key infrastructure hub for China's future robotics industry.
Beyond the factory tour, we attended a panel discussion featuring JD Industrial, Rokae Robotics, Huaqin Technology, and CASMT. These four roles represent four segments of the supply chain: platform, body manufacturer, ODM, and equipment provider.
The frontline reveals what lies beneath market hype. After half a day, I recorded eight insights.
01Scale First Emerges on the Component End
JD Industrial provided a set of data.
Over 500 companies procure robot components through the platform, including Unitree, AgiBot, Honor, and Huawei; there are more than 1,600 component suppliers, with over 300 added just last month; over 60,000 SKUs are actively selling; and procurement volume for components has increased by nearly 300% year-on-year.
Volume at the finished machine level is still emerging. On the component end, growth has already taken off.
A notable detail: about 30% of procurement by these 500 companies is completed "silently." No customer service inquiries, no price negotiations—direct orders placed, goods received, and assembled internally.
In essence, this reflects the current survival mode of labs and small teams, where even the inquiry step is skipped. They assemble the hardware themselves upon delivery. In a sense, this proves that the industry spark is still in gestation: orders are scattered, private, and bypass tenders.
JD itself is driving demand. Its group-level plan involves procuring 3 million robots, 1 million unmanned vehicles, and 100,000 drones over five years; the retail platform sets a sales target of exceeding CNY 10 billion annually for robots; over 300 finished machine brands have partnered with JD's self-operated channel. In terms of application scenarios, it possesses 20,000 offline stores, 700,000 logistics employees, and its own factory production lines, all generating real-world operational conditions daily.
A platform willing to act first as a user and then as a channel indicates it believes the bottleneck currently lies on the supply side.
02Component Makers Want Standards; Body Makers Fear Them
The head of JD Industrial's robotics business stated this bluntly:
"Component makers strongly desire standardization because it significantly improves efficiency and lowers costs. But body manufacturers strongly resist standardization, as once standardized, their value diminishes."
This statement essentially captures the internal contradiction within the supply chain today.
Component factories earn from shipment volume; the more unified the specifications, the less they spend on molds, and the lower the unit cost. Body manufacturers earn from differentiation; the more unified the standards, the easier they are to replace by peers, and the weaker their bargaining power.
Thus, the word "cost reduction" contains inherent contradictions between the upstream and downstream halves of the robotics industry.
This also explains why industry standards have been slow to advance despite years of calls: those truly pushing them are platforms and component/equipment manufacturers; those with the loudest voices often spend most of their time showing off.
03Domestic Substitution Has Won Volume, Not Yet Price
Song Bin, Vice President of Rokae Robotics, cited two figures: global industrial robot shipments are approximately 650,000 units per year, with the Chinese market accounting for 350,000 units, over half of the total; the market share of domestic brands has risen from 15% in 2015 to approximately 55% in 2025.
Ten years, from 15% to 55%. This is the most familiar growth curve of Chinese manufacturing.
However, the substantive quality must be clearly seen. In the four major processes, domestic players are still catching up. True overtaking occurred in collaborative robots—a category where overseas competitors have largely reduced to Universal Robots alone, while the rest are Chinese companies.
Standard product tracks win by market share; new category tracks win by positioning. Only the latter holds pricing power.
Rokae is a key player in this track. Listed on the HKEX in July under the banner of "the first fully-series intelligent robot stock," its industrial product line covers payloads from 3 kg to 300 kg, with arm spans ranging from 475 mm to 3.2 m; its collaborative line ranges from 3 kg to 45 kg. Globally, it has deployed nearly 50,000 units serving over 1,000 customers. It ranks first domestically in light-payload flexible collaborative robots and third domestically/seventh globally in multi-joint categories.
Certification is another hurdle. Rokae holds China's first national-level MTBF 80,000-hour zero-failure certification, China's first and the world's second IEC certification for medical surgical use, along with CE and functional safety certifications.
Its success stems fundamentally from the fact that industrial customers do not listen to stories; they only recognize the ability to solve practical problems.
04The Real Bottleneck Is Data, Which Has Never Been Collected Before
Song shared a technical detail: using Gripper A, Gripper B, or Gripper C to run the same program requires no repeated programming.
To laymen, this sounds mundane; to experts, the significance is profound. When tools change, the position of the robot's end-effector changes, requiring trajectory rewriting for the same motion. Now, it can adapt automatically, relying on spatial models.
The next layer down is data. His exact words were: "In human history, physical manipulation data has never been collected directionally."
Language models can capture decades of internet text, but there is no such internet for robot operation data. Polishing, gluing, screwing, handling, and welding in factories are all skills held by craftsmen's hands, never recorded as trainable samples.
So now two types are collected: data gathered on real machines, and data simulated (Sim2Real). Both are timestamped, aligned across force, tactile, and visual inputs, corrected, and fed into world models.
The market implication of this technical discussion is singular: whoever owns the scenario owns the data.
JD has 20,000 stores, 700,000 logistics employees, and its own factories; Rokae has nearly 50,000 operating devices and over 1,000 customers; Huaqin has production lines assembling 200 million products annually.
What were called fixed assets a few years ago are now called "mines."
05The ChatGPT Moment Will Likely Arrive in Factories Before Living Rooms
The mainstream narrative suggests that general-purpose humanoid robots entering homes to serve tea, pour water, and care for the elderly will mark the ChatGPT moment.
What I heard on the frontline contradicts this.
Take elder care as an example: models must handle real-time changing environments while addressing safety and ethics—a long road ahead. Industrial applications are the easiest to land: factory environments are standardized, tasks are fixed, and the requirement is to learn variations within specific processes, not to know everything.
Omniscient models are useless here. What is needed is a model sufficient for one industry, one edge device.
This judgment deserves attention. While the mainstream narrative bets on general models, frontline workers focus their efforts on edge solutions.
Rokae's client list speaks volumes: Schaeffler uses its wheeled humanoid robots for joint module automation lines; Xiaomi's dark factories employ its industrial robots; Foton Cummins has deployed over 100 sets; Honglu Steel Structure uses it for steel structure welding, reducing processing costs to two-thirds of industry peers; Bosch, ZTE, SERES Power...
Factories have no audience, only the real demands of Chinese manufacturing. Whether it works and how well it works can be determined in a single day.
06At Present, Factories Prefer Wheels Over Humanoids
In the factory, Song made a simple remark: Throughout history, how many skilled craftsmen work with their feet?
Factory floors are flat; craftsmanship relies on the coordination of hand, eye, and brain. Bipedal locomotion increases both cost and instability in industrial settings, making the ROI unjustifiable.
Rokae's path consists of three stages: fixed installation (fixed base); composite robots (AGV chassis plus robotic arm); and finally, wheeled dual-arm robots.
Composite robots have already been deployed at Goertek and Foxconn sites. One unit can manage 10 machine tools simultaneously, shifting efficiency from 1-to-1 to 1-to-10.
Wheeled dual-arm robots take a further step forward: equipped with cameras in the head, they can perform both handling and quality inspection.
The calculation factories make when buying robots is based on the magnitude of efficiency gains. Whether it looks like a human currently does not affect this bill.
07The Logic of Building Robots Is Opposite to Building Smartphones
When Mr. Wu from Huaqin said this, I initially didn't take it seriously, but understood after listening further.
SMT automation rates on smartphone lines can exceed 95%, assembly lines over 70%, with almost no people along a 300-meter walk. Smartphones involve large-scale ramp-up after finalization, with tens of millions of units per model.
Robots are different: each unit varies; clients may order only2, 3, or at most dozens to hundreds of units.
Therefore, a 300-meter smartphone line is split into nearly 20 Cell lines. Each Cell installs a module, tests it, and then sends it to the main assembly line.
In terms of investment, this approach is expensive at present but necessary. The scale logic of consumer electronics cannot yet be transferred; what can be reused are quality systems: MES, data traceability, and test cases are reusable; capacity and takt times are not.
Testing differs too. Consumer electronics undergo functional testing in fixed environments; robots require test environments configured according to client sites. Whether bipedal or wheeled dual-arm, different usage scenarios dictate different aging test cases.
Currently, many consumer electronics companies view robots as a new growth curve, betting that "stability in mass production" will eventually become a barrier to entry.
08Three Unresolved Knots: Cost, Standards, and Going Global
The following frontline insights come from Dr. Yan of CASMT.
The first knot constraining the robotics industry lies in process engineering.
For example, ball screws currently use grinding; precision can be achieved, but costs remain high. Domestic suppliers providing to international leaders report that "quoted costs may not even cover expenses."
The solution involves switching from grinding to turning, or using laser heat treatment to control deformation. CASMT built the world's first fully automated roller screw line for this purpose: 30-second cycle time; hollow cup motors at 25 seconds; joint modules at 30 seconds; frameless torque motors at 50 seconds, reducible to 30 seconds if needed.
The second knot lies before the convergence period.
Dexterous hands currently have four technical routes—linkage, tendon-driven, four-cylinder, and small joint modules—none of which have converged, with prices reaching CNY 200,000 to 300,000 per pair. Six-axis force sensors contain too many strain gauge wires for automated assembly, limiting operations to test benches.
There are currently no mandatory standards for robots. Consequently, the industry borrows automotive TS16949 and Failure Mode and Effects Analysis (FMEA) frameworks for quality management.
An industry needing to borrow standards from another to manage quality speaks volumes about its maturity.
The third knot lies in service radius.
A standalone arm manufacturer cannot succeed alone; they must collaborate and ride larger ships. Overseas markets demand lawn mowing robots, public cleaning robots, and performance robots—places with large yards, high labor costs, and willingness to pay for labor savings. Domestic scenarios are entirely different.
This explains why JD is building 80 ROBOBASE industrial parks and training 100,000 after-sales engineers. For robots going global, half the product sold is hardware, and the other half is the promise that "someone will fix it if it breaks."
09Conclusion
The phrase "China's next economic engine" has been frequently mentioned this year. From the frontline perspective, this engine has ignited, but the transmission shaft is still breaking in.
China possesses the most complete range of production categories globally, which will be our greatest asset in this round of robotics industry revolution.
But when the RPM will rise depends on whether that 30 seconds in the workshop can be compressed to 20. Who stands in the center position at launch events tells us little.
How many processes an embodied arm goes through from parts to line-off is difficult for ordinary people to grasp, even after visiting.
30 seconds, 25 seconds, 50 seconds. These numbers running in frontline factories are hard to narrate and unsuitable for PPTs, yet they are the true tachometer readings of this engine.
Every tick on the tachometer represents the real progress of China's robotics industry.