Summary

Yifei reaffirmed in 2023 that it is an industrial robotics company: industrial robotics rose from 10% of total revenue to 20% in 2024 and may increase further this year. Non-standard integration wins orders but makes scale and headcount grow linearly; standardized products expand margins through mass production, falling costs and iterative feedback. The two businesses therefore had to be separated in staffing, organization and office locations. “At our core, we are an industrial robotics company, not a systems integrator.”
Zhang Sai is betting on embodied intelligence not to redo box-moving and screw-tightening with humanoid robots, but to crack flexible applications where traditional automation cannot make the economics work. High-mix, low-volume production, frequent changeovers, flexible assembly and textile handling are not absolutely impossible for industrial robots; the problem is that non-standard investment can lose its value in less than 6 months. Foundation models, deep learning and generalized training now appear to offer the first hazy path toward solving needs that could long only be met with “I can’t make it work.”
Yifei believes the core advantages of industrial embodied intelligence are real demand, customer trust and factory data—not impressive demonstrations. Zhang Sai says box-moving may simply be easy to show off technically, while its wet-wipe lid applicator already runs at 120 packs per minute and replaces 4 workers; replacing it with 4 human-shaped robots would be “a step backward.” Yifei’s more valuable assets are its thousands of cases, customer access to factory environments and training data available only after years of cooperation.
Zhang Sai’s deployment cadence is “release a demo within the year, begin real-world deployment the following year,” but broader adoption still depends on a capability leap that has yet to occur. Teleoperation, visual training and simulation training are being explored in parallel. If a moment of collective evolution arrives in the next 2 years—“once one robot learns, every robot knows”—he believes embodied intelligence could begin spreading over the following 3–5 years.
The effective path for Chinese robotics companies is to turn every custom project into reusable modules, then productize the robot itself. Yifei further breaks loading and unloading, glue application and packaging into modules such as tray loading, tray transfer and tray stacking; roughly 60%–70% of a new project can be covered by accumulated know-how. On the standardized-product side, it produces in batches of 50 or 100 units and triggers replenishment when inventory falls below a preset safety level, shifting from project economics to product economics.
Going global is first a margin equation designed to escape domestic price competition, and only then a brand story. Zhang Sai says there should be data showing domestic brands sold more industrial robots in China last year than imported brands, and believes Chinese companies can still win on the “Big Four’s” home turf. Yifei follows customers such as Lens Technology, Foxconn and GoerTek into Vietnam, Thailand, India and Mexico, while turning scattered overseas buyers into distributors and local service hubs.
Yifei’s 13-year history shows that the technology moat in industrial To B must be delivered alongside founder-led sales and on-site service. The company grew from 3 people in an 8-square-meter office to a 7-person team that stayed with its first robot until “breaking every day” became “breaking once a week.” Zhang Sai forced himself to become the company’s “biggest sales”; his standard is to make customers either “bowled over by your technology or won over by your passion,” with the final test still being a signed order and successful delivery.

Deep dive

1. China’s Industrial Robotics Battleground Has Shifted From Import Substitution to Productization and Global Competition

The opening industry backdrop: industrial robots were born in the United States, but Japan and Europe were the first to pay premium prices for them after the war, facing labor shortages and an urgent need to restore precision manufacturing. That was how the “Big Four” rose. China became the world’s largest industrial robotics market roughly 12 years ago, and later became the largest market on both the production and sales sides.
Yifei’s 13 years map almost perfectly onto this transition. The company rode China’s precision-manufacturing upgrade while absorbing the manufacturing sector’s price-war pressure, starting with parallel robots and gradually moving into production lines for consumer electronics, food, biopharma, personal care and finance.
The company’s history was not a straight climb. Li Feng recalled that roughly 10 years ago he invested in an industrial robotics startup that was visibly insolvent, whose founders fell out before the company eventually broke apart. Zhang Sai later climbed out of that trough, and the company is now preparing to list on the Hong Kong Stock Exchange. Li sees that trajectory as evidence of Zhang’s resilience and ability to solve problems.

2. The 2023 Transformation Put the Robot Body Back at the Center of the Company

Zhang Sai describes Yifei since 2023 as having undergone a “transformation.” The company was one of China’s earlier parallel-robot makers, but the market eventually remembered it mainly for turnkey solutions. “Many people had already started not knowing that Yifei made parallel robots,” which Zhang considered the most dangerous form of brand drift.
Before that, customers could rarely buy a Yifei robot on its own. What Lens Technology and GoerTek typically received on-site was a complete system. Zhang revisited the company’s core identity and concluded: “At our core, we are still an industrial robotics company, not a systems integrator.”
Yifei therefore spun its industrial robotics business into a subsidiary called Yifei Robotics, while “Yifei Automation” became a sub-brand. Industrial robotics rose from 10% of total revenue in 2023 to 20% in 2024, and Zhang said the share could rise further this year.
In 2023, the company completed its shareholding reform and formally launched an IPO project called “AllSpark.” Zhang named it after the “AllSpark” in Transformers, describing the spark as a guide for the company’s future development. The unchanged ambition is “to build a world-class Chinese robotics company”; the prospectus puts it as “using industrial robots to lead humanity into the intelligent age.”

3. Industrial Robots Are Already Embedded in Everyday Consumer Products—End Users Just Cannot See Them

Food is the clearest proof of operating life. Yifei began working with Dong’e Ejiao’s Taohua Ji packaging line in 2016, which went into production in 2017. More than 50 Yifei robots have run continuously for roughly 8 years, 24 hours a day. Similar equipment sits behind Xinghualou mooncakes, small cakes and snacks. A large North American bakery also opened Zhang’s eyes to product categories such as ciabatta, focaccia and croissant.
Consumer electronics is Yifei’s largest business. Lens Technology alone has deployed thousands of equipment sets; a piece of phone glass passes through more than 30 processes from substrate to finished product. Yifei has also participated in the assembly of Sony PS5 and Nintendo Switch controllers at Foxconn, GoerTek and other factories.
The application range also includes cosmetics lines at Bloomage Biotech, cash sorting and allocation in the vaults of Agricultural Bank of China and China Construction Bank, and wet-wipe packaging. Zhang’s everyday prompt is: “Take a pack of wet wipes from around you… that lid was glued on by our robot.”

4. The First Robot Was Iterated Into Existence at Qidu Pharmaceutical

The company began with 3 people, an 8-square-meter office and 2 desks. It took the team roughly a year to build a robot from scratch. No customer was willing to buy the first unit, so Zhang leaned on his Tsinghua and Columbia background to secure free trials: “Buy it after it works well for you.”
Shandong Qidu Pharmaceutical became one of the earliest trial customers. The company had only 7 people at the time; the others had no cars, so nearly everyone squeezed into Zhang’s small car for trips between Jinan and Linyi. After the equipment was unpacked, workers proactively gathered around to install it. The caption on the photo read: “The people of Qidu are too warm-hearted—we couldn’t even get our hands in.”
The first unit initially “broke every day.” The team stayed on-site and kept improving it; the interval between failures gradually stretched to 2 days, then 1 week, before the robot finally no longer required a permanent presence. In 2013, they used a crude solution: multiple dashcams recorded continuously, and the line had to stop whenever a failure occurred so the footage could preserve diagnostic clues.
The Chery A1 is still at the company, with its scrapes and peeling paint left untouched as a physical reminder to “never forget the original mission.” In 2018, Zhang switched to a BYD Qin for an equally direct reason: “I run a Chinese industrial robotics company, so I first have to believe in Chinese products.”

5. 3C Drives Automation Through Iteration Speed; Everyday Consumer Goods Pay It Back Through Long Product Lives

Zhang calls 3C electronics China’s second-largest industrial robotics market after automobiles. It started later but grew faster: industry participants were more willing to accept new things, products iterated frequently and capital investment appetite was strong. The sector therefore naturally became Yifei’s largest customer base.
Food, pharmaceuticals and personal care operate differently. Their individual products have long life cycles, allowing automation investment to be amortized over years of stable production. The Taohua Ji packaging line has not been replaced in 8 years. A single Danhong injection product can generate more than RMB2B in annual sales, allowing the company to keep increasing UPH and replicate the line a second and third time.
After several major events, customers’ understanding of automation moved up “a level.” Factories that once thought they could “just have people cover for it” later discovered that, under extreme conditions, people could not cover the gap at all—or could not be used. Seeing lights-out factories continue operating without anyone inside prompted companies to keep upgrading their automation.

6. The Unilever Breakthrough Came From Replicating a Demo Line Next to the Customer

When Yifei first made contact in 2015 and began formal discussions in 2016, Unilever accepted the automation concept but did not trust Chinese robots. The opening was a sauce product at its Jinshan plant: the oval bottle did not suit conventional sorting equipment, and end-of-line packaging still depended heavily on manual labor.
Yifei rented a site directly beside the Jinshan factory and built a demo system around the customer’s conveyor line. Because there was no upstream filling machine, Zhang hired several cleaners to keep placing bottles, manually simulating the real line speed, then invited the plant manager to see whether the robot could load the bottles.
The Jinshan plant agreed to trial the system and achieved good results. The case was reported to the China head and later reviewed by the global head of manufacturing. Only after that trust was established did Yifei receive more orders, expanding into Knorr chicken essence, monosodium glutamate and concentrated soup products, as well as Magnum, Cornetto and milkshake cups under Wall’s.
The case captures the early foothold for Chinese vendors: not “one robot conquers every application,” but identifying a non-standard gap that conventional equipment cannot handle and using a visible, functioning demo to build trust from zero to one.

7. Non-Standard Work Can Open an Industry, but Without Reusable Know-How It Cannot Scale

Zhang’s view of a pure non-standard model is blunt: “You can keep doing non-standard work forever, but I don’t think you can build a large company that way.” Re-customizing every project makes revenue and headcount grow linearly. He has seen non-standard companies become large only by accumulating huge workforces, and says: “Just thinking about it gives me a headache.”
Yifei did not reject non-standard work; it treated it as raw material for product extraction. The company first abstracted scenarios such as loading and unloading, glue application and packaging, then broke them into functional modules such as tray loading, tray transfer and tray stacking, turning one-off delivery into reusable accumulation.
For new scenarios, the team works back from more than 1,000 existing cases. Even when new development is still required, roughly 60%–70% of the workload can be covered by historical accumulation. More importantly, experienced engineers can now look at a requirement and quickly judge whether “this thing can be done.”

8. Standardized and Non-Standard Products Are Not Two Product Categories but Two Opposing Operating Systems

Non-standard automation is a customer-based project business: production, procurement, inventory, quality inspection and shipment all follow orders, while the team accumulates know-how around major customers. Lens Technology has roughly 5 or 6 factories in Changsha, with additional sites in Xiangtan, Songshan Lake and Vietnam, yet Yifei still serves them across factories with the same customer-familiar team.
The robot body is product-based, or “product plus location-based.” The same parallel robot can handle 3C loading and unloading or food packaging. The core requirements are ease of use, durability and portability—not endless reshaping for a single customer.
The standardized-product flywheel is to expand output, lower costs, improve margins and use feedback from more customers to keep improving the product. Yifei can produce 50 or 100 units at a time and distribute them as orders arrive. It can, for example, set a safety-stock level and trigger production of the next 100 units once inventory falls below it.
Zhang therefore refuses to let the same group handle both businesses: “I don’t think this is workable. It will definitely lead to mistakes.” Yifei ultimately separated the personnel, organizational structures and even physical office locations, allowing project-based work and batch manufacturing to run at their own pace.

9. Technical Founders Cannot Use “Introversion” to Opt Out of Selling

Zhang started out as an engineer who drew designs, but believes the top executive must reset expectations from the first day of the company: they cannot keep emphasizing that they are a “technical nerd” or “not good at speaking” and use that as an excuse to avoid sales. If they are unwilling to take on the responsibility, they should not be the company’s founder-CEO.
Sales communication can be trained. Practice in front of a mirror or with family, record it, listen back and refine it. A technical founder’s advantage is understanding the product’s underlying value better than anyone else. The task is to communicate it with enough force that customers are either “bowled over by your technology or won over by your passion,” with the ultimate purpose still being to close the deal.
Zhang often calls himself the company’s “biggest sales.” Even when the team has stronger salespeople, the founder cannot remain absent forever when a deal requires founder-to-founder communication, a display of technical strength or a signal of confidence in the company.
He also admits that the founder sometimes appears merely as a “mascot.” The key is to perform that role well and become the sales team’s backing. At the same time, the company still needs people who genuinely understand sales to stay close to customers and handle their demands, combining founder positioning, individual ability and a professional team.

10. What Embodied Intelligence Really Needs to Replace Is “Human-Powered Automation,” Not Mature Machines

Zhang stresses that Yifei is not a company built to serve VCs or government buyers, nor one chasing the latest trend. After 13 consecutive years in industrial robotics, it entered embodied intelligence from customer needs that remain unresolved on the factory floor—not from the hottest form factor in the capital markets.
Traditional robots have replaced large amounts of repetitive, single-purpose labor but struggle to enter flexible, generalized, frequently changing, high-mix, low-volume scenarios. The issue is usually not that the task is technically impossible; it is that the solution costs too much and the return on investment does not work, leaving people as the more cost-effective option.
A typical site is “human-powered automation”: workers line both sides of a conveyor, with one person tightening a screw and another pressing in a component. Industrial robots could handle the work, but the customer may switch products as soon as that order ends. A custom system could lose its value in less than 6 months, turning the initial investment into a write-off.
For stable products that do not change for 7 or 8 years, traditional industrial robots still have a decisive advantage: they are fast, precise, stable and tireless. Embodied intelligence is not targeting the market already suited to conventional automation; it is targeting the blank space where changeover costs and insufficient flexibility have kept traditional automation from penetrating economically.

11. Flexible Assembly and Textile Handling Have a Hazy Window of Opportunity Thanks to Foundation Models and Other Technologies

Flexible assembly depends on human touch and real-time adjustment. A part may need to be “shaken, tapped and rotated” before it fits. Traditional robots are too precise: when they encounter slight deformation, they continue pushing along a fixed trajectory and may even crush the part.
Textiles are more extreme. A factory can automate material delivery, but fabric is too soft for robotic hands to grasp reliably. Sewing still depends on large numbers of skilled workers. Only the tactile feedback and coordinated movement of 5 fingers can continuously handle wrinkles and deformation.
In the past, when customers asked for such applications, Zhang could only say: “Sorry, I can’t make it work. It’s too difficult.” Foundation models, deep learning and generalized training did not solve the problem immediately, but for the first time made him feel that “a solution seems to be appearing.”
When ALOHA released videos of tasks such as zipping up a jacket and putting on a pillowcase, Zhang still found them “highly disruptive,” even though they relied on teleoperation and extensive training. Traditional industrial robots could not previously perform such flexible operations; training the actions at all was already a major achievement.

12. Yifei Refuses to Move Backward Into Box-Moving and Screw-Tightening Just for the Sake of a Demo

Yifei’s training agenda initially came entirely from existing customers’ industrial scenarios that “could not be done before.” It did not prioritize pouring tea, folding clothes or cooking. Zhang’s commercial loop is straightforward: once the training succeeds, the company can enter new areas that its existing industrial robots cannot cover.
On the box-moving and screw-tightening demos shown by many robotics companies, Zhang’s view is that “they may not be real needs.” They may simply be easy to complete and therefore suitable for technical demonstrations. Industrial robots already perform these tasks faster; there is no reason to redo them in another form.
The wet-wipe lid applicator is his counterexample. The equipment can handle 120 packs per minute and replace 4 workers. If a mature machine were discarded and 4 human-shaped robots placed there to do the same job, “I think that would be a step backward.”
This does not mean Yifei cannot begin training with simple tasks. It means embodied intelligence must create incremental value on tasks that conventional equipment cannot complete economically before it qualifies as real demand.

13. Industrial Customer Relationships Turn Factory Access and Data Into Critical Embodied-Intelligence Assets

Zhang sees Yifei’s first advantage over academic or pure-model teams as: “We understand what the customer actually needs.” The second is the trust built through long-term service. Customers are willing to open their factories so the team can test, collect data and complete demos.
Industrial training volumes cannot match civilian scenarios not only because of algorithms, but because factory data is highly confidential and usually cannot leave the customer’s premises. Without commercial relationships and delivery credibility, a robotics team may not even enter the site, let alone obtain the relevant training data.
Yifei’s large base of cases and customer relationships also helps the team determine which tasks already have faster, more suitable dedicated machines and which steps still rely on people because of flexibility and changeover requirements, directing R&D resources toward real demand.
Zhang retains the uncertainty: “This is very difficult—genuinely difficult.” Customers and data only mean there is “a possibility of success”; they do not amount to ready-made model capabilities or guaranteed deployment.

14. Near-Term Embodied-Intelligence Delivery Has a Plan; Adoption Speed Depends on Collective Evolution

To pursue a technical path distinct from traditional industrial control, Yifei hired a dedicated embodied-intelligence team rather than forcing its legacy industrial team to transform overnight. The latter provides scenarios, robot bodies and engineering support.
Zhang’s company-level timetable is to release a demo within the year and begin deployment in real-world scenarios the following year. The judgment is grounded in existing customer demand and factory access, but a single-point demonstration is not the same as scaled commercialization.
His ideal end state is simple: “Originally, people did the work. The machine learns exactly, step for step, how people do it.” Achieving that still requires extensive training, while teleoperation, visual training and virtual-environment training have yet to converge on one winning route.
The longer-term variable is the possibility that AI capabilities suddenly leap forward, as language models did. The advantage of silicon-based life is that “once one robot learns, every robot knows.” If such collective-intelligence evolution appears in the next 2 years, Zhang estimates embodied intelligence could begin spreading over the following 3–5 years; otherwise, the timetable will have to move back.

15. Overseas Profit Comes From Escaping Price Wars; Scale Comes From Customer Migration and Local Service Networks

Zhang puts the underlying logic of going global bluntly: “It’s still because competition at home is too intense.” Overseas, companies can avoid the domestic price war, compete with foreign brands and earn better margins.
He believes there should be data showing that Chinese industrial robot sales exceeded imported-brand sales in China last year. In his view, that means the “Big Four” can no longer beat domestic companies in China. If Chinese companies enter the Big Four’s home markets while overseas brands remain attached to their “arrogant posture,” Chinese vendors may still win.
The first route is to follow major customers overseas. Lens Technology, Foxconn and GoerTek have expanded factories into Vietnam, Thailand, India and Mexico, where there are relatively few local automation vendors or systems integrators to serve them; they therefore bring Chinese suppliers along. Overseas factories built by new-energy customers such as CATL offer similar opportunities.
When overseas entities place orders directly, some circulation, tariff and transportation costs can be eliminated, allowing Yifei to charge more than it does in China. These orders already come with an established base of trust, making them an easier first step than entering a completely unfamiliar market.

16. Overseas Acceptance Has Improved, but After-Sales Capability Determines Whether Orders Keep Expanding

The second route overseas is to build a distribution network. Yifei first acquires scattered customers through overseas trade shows, Facebook, LinkedIn, Google SEO and Alibaba International, then persuades buyers with good user experiences to become distributors. One-off purchases become recurring referrals and profit sharing. The network now covers Southeast Asia, Europe, North America and South America.
Zhang recalls that overseas customers were still reluctant to accept Chinese robots 5 or 6 years ago. Some early Chinese robotics companies acquired family businesses that had operated for 200 or 300 years, mainly to obtain a name and brand, then gradually changed perceptions by putting “new wine in old bottles.” Other technology companies adopted dual brands at home and abroad. By the time Yifei began going global, earlier players had already completed a significant portion of the market education.
Localization begins with organizational capability. Yifei created a dedicated overseas team, including members who speak English or Japanese and can travel abroad and serve customers. The company also helps with visas and APEC cards. Products are tuned in China toward “idiot-proof operation”; some customers in Italy and Brazil have already brought systems online using only manuals and video guidance.
The ultimate goal is not to keep flying Chinese engineers around indefinitely, but to turn customers into local service nodes. Yifei has an office and demo room at an Italian distributor’s site and trains its engineers to handle on-site service. Overseas service can be charged separately, giving the distributor a revenue stream while allowing Yifei to expand its overseas share and grow alongside the distributor.

The Secret Engine of AI - Prolific [Sponsored] (Sara Saab, Enzo Blindow)Chinese AI – They're Just Like Us? With Beijing-Based Concordia AI CEO Brian Tse