Chapter 2: Selecting Leaders — Capital, Revenue, and Visibility

Overview

The first problem in mapping Chinese robotics companies is not the number of firms. It is the unit of comparison. Funding raised by a private humanoid startup, consolidated revenue reported by a listed automation group, new orders booked by a warehouse-robot supplier, cumulative shipments claimed by a service-robot vendor, a tactile startup’s product specification, and a count of edited news articles describe different events. Putting them into one “scale” column would create a neat table and a misleading conclusion.

This atlas therefore locks exactly 24 companies for deep profiles and 20 for atlas-only coverage. “Deep” is a commitment about the amount of research and explanation allocated in later chapters. It is not a performance grade, a top-24 ranking, or a prediction of investment return. Atlas-only companies remain necessary to represent the breadth of the ecosystem, but they receive more compact treatment because evidence is thinner for some profile fields, their role overlaps with a deep company, or their main purpose is to provide a comparison point.

Blank cells are part of the result. Undisclosed revenue or shipments are recorded as undisclosed, not zero. A small set of search results does not establish zero media attention. Media visibility can be counted only after the fixed interval, outlet set, body mentions, name collisions, and syndicated copies have been reviewed. That review is not complete, so visibility remains unmeasured for every company in this edition.

After reading this chapter... - You can explain why the 24/20 split describes research depth rather than company rank. - You can keep funding, valuation, revenue, orders, production, shipments, installed base, customers, and media visibility in separate fields. - You can distinguish a measured zero from an undisclosed, unmeasured, inapplicable, or unresolved value. - You can reproduce the logic of the fixed media corpus covering August 17, 2025 through August 17, 2026. - You can turn the atlas into verification questions for procurement, a manufacturing-cell trial, or a WRC visit.

2.1 Selection builds a research map, not a podium

The selection question is not “Which Chinese robotics company is best?” That question omits the application, time horizon, risk boundary, and buyer. A regulated surgical system, a humanoid used for policy research, an autonomous mobile robot repeating warehouse moves, and a tactile sensor measuring fingertip contact do not share one utility function. The 44-company roster is consequently a sampling design for investigating a value chain, not a tournament bracket.

Six signals shaped coverage. The first is resolvable legal and brand identity. The second is capital evidence with an explicit metric type. The third is audited revenue or bounded operating traction. The fourth is technical and ecosystem evidence, including product documentation, developer resources, primary research, and integration scope. The fifth is visibility in a fixed editorial corpus. The sixth is relevance to questions that a WRC 2026 visitor can test. No one signal is a pass-fail gate, and admission does not imply that all six are complete.

The method deliberately avoids a weighted composite score. Giving the same capital weight to an early-stage startup and the same revenue weight to a listed industrial group erases lifecycle differences. Giving media coverage a large weight privileges consumer brands, spectacular demonstrations, and firms active in capital markets over quiet component suppliers. Making a confirmed WRC booth mandatory would omit consequential firms that do not exhibit or whose official records remain unsettled.

The deep label is also not evidence of superiority. A company may deserve deep treatment because it connects several chapters, represents an otherwise missing layer, exposes a useful uncertainty, or provides unusually rich source material. An atlas-only company may be highly consequential but need fewer pages to prevent repetition. Readers should neither infer ranks one through 24 inside the deep set nor interpret the other 20 as companies that failed selection.

2.2 Separate the legal entity, brand, listed vehicle, and product

Before comparing companies, the atlas must establish what it counted as a company. Assuming that one English brand maps neatly to one Chinese registered entity can attach parent-company revenue to a subsidiary or convert product-line activity into legal-entity performance. Counting Chinese names, English names, and abbreviations as different firms can duplicate media articles and funding rounds. The data model therefore keeps brand, registered entity, listed vehicle, parent or subsidiary relationship, and product line distinct.

The National Enterprise Credit Information Publicity System is the starting point for Chinese registration checks. It is an authoritative registry, although dynamic queries and interactive verification can make a specific result difficult to reproduce as a simple static link. More importantly, registration proves neither product performance nor adoption ^[1]. When registry evidence is insufficient, the process uses exchange filings, regulator records, official developer material, and multiple independent records. An unresolved identity stays unresolved.

Evidence strength depends on whether a source is fit for the claim. An audited annual report is strong for a defined accounting line in a defined reporting period, but it does not establish the field success rate of one robot. Developer documentation can establish interfaces and supported products, but not independent durability or customer satisfaction. A registry resolves an entity, not a valuation. A primary paper can test a method on specified robots and tasks, but it does not substitute for company revenue or production evidence.

2.3 The locked 24 deep profiles

The next table contains all 24 deep-profile companies. “Primary lens” is a question to open later chapters, not a permanent single-industry classification. Many firms span humanoids, mobile manipulation, data, software, and components. The links are starting points for identity or product-scope checks; no link verifies every business claim about that company.

The last column is an interpretation contract, not a list of defects. An official product page establishes that the issuer presents a product and support scope; it does not prove that every model is at the same production stage. Even an exchange report requires the reader to separate consolidated finance from a robotics segment, new orders from recognized revenue, and closing backlog from contracts signed during the period.

2.4 The locked 20 atlas-only companies

The atlas-only roster preserves breadth around the deep set. These companies connect humanoids, collaborative and industrial arms, end effectors, mobile bases, service robots, consumer products, and medical systems. They serve as technical or commercialization comparators in later chapters without receiving the full profile length.

Together, the two tables contain exactly 44 named companies. Adding or removing a firm might improve a future sampling design, but that must be an explicit revision. Quietly adding a fashionable name or dropping a difficult identity would shift later comparisons and the field route. New candidates can be tracked separately without altering the locked 24/20 contract for this edition.

2.5 The grammar of numbers: capital is not operation

The most common comparison error is sorting numbers whose nouns differ. This atlas keeps funding, valuation, revenue, orders, production, shipments, installed base, customers, and media visibility in separate columns. Each metric answers a narrower question.

LimX announced a US$200 million Series B in February 2026 and a nearly US$200 million pre-IPO round in July, and the latter release states that the company raised US$400 million in total over the previous six months ^[3]. The rounds can therefore be kept distinct at issuer-disclosure level, while retaining the “nearly” rounding and company-attribution caveats.

Geek+ reported RMB4,137 million in new signed orders for 2025. That is a disclosed contract-flow measure, not revenue for the same period ^[4]. Orders may inform future demand, but delivery schedules, cancellations, and recognition policies remain relevant. Rewriting “RMB4,137 million in new orders” as “RMB4,137 million in revenue” would merge two accounting events.

Pudu’s 2025 brochure reports, as of March 2025, more than 90,000 robots shipped, coverage in more than 60 countries and regions and more than 600 cities, and a network of more than 700 global distributors ^[5]. The statement is useful when its date and issuer qualification remain attached. Shipments, however, are not the current active installed base; country count is not paid-customer count; and a company brochure is not an audited report. The appropriate field is “issuer-reported cumulative shipments,” not “independently verified active robots.”

The missing cases are equally important. Deep Robotics provides product brochures but no finance in the cited source ^[6]. DexForce documentation resolves W1 and software scope but not finance, shipments, or deployment success ^[7]. Noitom describes data-factory and retargeting work but does not disclose finance on the cited company page ^[8]. The correct entry is undisclosed, not zero. Zero is an observed quantity; undisclosed describes the limits of observation.

That distinction also prevents a disclosure bias. A listed company that reports many figures should not look weaker merely because failures and costs are visible. A private company should not be assigned an invented floor merely because it reports selectively. The atlas evaluates both the value and the disclosure status, keeping uncertainty visible rather than converting it into a pseudo-measurement.

2.6 Media visibility: a bounded one-year corpus

Media visibility is not an impressionistic fame column. The measurement interval is August 17, 2025 at 00:00 through August 17, 2026 at 23:59:59, China Standard Time. Fixed start and end points prevent company searches conducted on different days from silently receiving different windows. A future edition can move the window while retaining the same rules.

The fixed outlet set is Reuters, Bloomberg, Financial Times, Nikkei Asia, South China Morning Post, TechCrunch, Caixin, Yicai, 21st Century Business Herald, Securities Times, China Securities Journal, The Paper, Jiemian, 36Kr, TMTPost, LatePost, People’s Daily, and Xinhua. It combines international English-language outlets with Chinese English and Chinese-language business, technology, and general news. An article outside the set is not deemed inferior; it simply does not contribute to this bounded measure.

Queries use the resolved legal name, official English and Chinese names, and verified aliases. GDELT DOC discovery is combined with publisher archives and reproducible site queries. A result is not counted at retrieval time. The article body must refer to the resolved company. Press releases, sponsored posts, job listings, directories, quote-only pages, social posts, and name collisions are excluded. Near-identical syndicated copies are clustered using title, lead, wire source, and date ^[2].

Collision review is essential. Short brands and ordinary words can retrieve unrelated entities in China or elsewhere. Chinese articles may omit the English brand or mention only a subsidiary, creating false negatives. Paywalls, indexing delay, language segmentation, and removed pages introduce further bias. Search-engine hit estimates are therefore never used.

The intended outputs are the number of eligible articles and the number of distinct outlets. Within-roster quartiles and a measured zero can be reported after the review, but zero is allowed only after every verified alias and collision has been checked. Retrieval routes exist for GDELT, Reuters, South China Morning Post, Xinhua, and the other outlets, yet article-level export, body confirmation, entity disambiguation, and wire deduplication are incomplete ^{[10] [11] [12]}. All 44 media values consequently remain unmeasured and unpublished in this edition.

Even a completed visibility count is not a commercialization score. Listings, consumer familiarity, dramatic demonstrations, and large funding rounds can increase coverage. A component supplier operating under customer confidentiality may receive little editorial attention despite considerable operational influence. The metric answers “How visible was the resolved company in this corpus?” It does not answer “Whose robot works most reliably?”

2.7 Missingness, counterevidence, and an unresolved identity

A transparent atlas distinguishes why a cell is blank. “Undisclosed” means located public sources did not provide the value. “Unmeasured” means a defined measurement process is incomplete. “Not applicable” means the metric does not fit the business form. “Unresolved” means the entity or attribution cannot yet be established. Rendering all four as a dash would hide the difference between a real zero and pending research.

SynapX is the clearest boundary case. PaXini’s official partner material names SynapX and shows the joint system. At the research cutoff, however, no authoritative registered entity identifier, independently resolved official domain, financing record, or independent performance metric had been established ^[9]. Partner confirmation is evidence of a relationship, not a registry record. The atlas neither removes SynapX silently nor promotes partner material into full verification. It retains SynapX in the deep roster while marking legal identity, official domain, and finance unresolved.

Counterevidence search is not a ritual for diminishing large claims. It is a method for finding definition conflicts. The process pairs funding with valuation, revenue with gross merchandise value or bookings, new orders with backlog, production with shipments, shipments with active installations, pilots with repeat paid deployment, and autonomy with teleoperation. When sources disagree, the first check is whether they refer to the same entity, date, unit, reporting scope, product, and test conditions.

Company sources have selection bias. They naturally emphasize successful demonstrations, strongest specifications, and representative customers while disclosing less about failures, returns, maintenance, and integration cost. Audited filings have clearer accounting boundaries but may lack product-level technical detail. News offers external observation but is shaped by access and newsworthiness. Academic work can offer controlled evaluation while remaining bounded by its embodiment, dataset, and task. No source class answers every question.

Evidence-confidence checkpoint

The “verified” label remains scoped to a claim. A verified legal identity does not upgrade every product claim. A verified annual report does not verify a demonstration. A partially verified official specification may still be useful for integration planning, provided it is tested rather than treated as comparative proof. Evidence confidence should travel with the cell, not become a permanent badge attached to the company.

2.8 Practical interpretation: narrowing candidates for one cell

Consider a manufacturing team evaluating a cell that picks irregular parts and inserts them into an inspection fixture. Sorting the 44-company roster by capital is not the first step. Decompose the task into localization, approach, grasp, contact-rich insertion, inspection logging, and failure recovery. That exposes different candidate roles: 3D perception from firms such as Mech-Mind; force-controlled arms from Flexiv or ROKAE; hands, touch, and end effectors from LinkerBot, PaXini, Tashan, DexRobot, Wuji, or DH-Robotics; data and learning layers associated with DexForce, AI², Noitom, or SynapX; and mobile orchestration represented by SEER.

The team then asks the same integration questions of each candidate. Does documentation expose the necessary coordinate frames and control rates? Can raw force or tactile values be logged? Are safety-stop and restart states visible to the cell controller? Will the vendor provide failure cases and maintenance intervals? Is the evaluated unit the same configuration as the sold unit? Are paid customers separated from pilots and ecosystem members? Total funding cannot substitute for the quality of those interfaces.

Operating claims also receive metric-specific questions. “Mass production” should be followed by period, completed-unit count, model mix, and acceptance-test definition. “Large shipment” requires paid versus free, period versus cumulative, and return treatment. “Hundreds of customers” requires separation of direct buyers, pilot sites, channel partners, and ecosystem registrations. “Revenue growth” requires entity scope and robot-product contribution. A missing answer is not an automatic rejection, but it becomes an uncertainty cost in procurement.

Finally, conduct a matched cell trial. Use the same part family, lighting, fixture tolerance, and success definition. Record not only nominal success but human interventions, recovery time, excessive contact force, recalibration frequency, and traceable logs. Do not combine one vendor’s best brochure specification with another firm’s audited revenue in a single score. Capital evidence informs continuity and investment capacity; revenue and orders inform traction; documentation informs integration; the cell test informs task performance.

Missingness becomes operationally useful in this workflow. Undisclosed finance does not mean the technology cannot pass a trial, but long-term supply risk needs separate diligence. An undisclosed installed base does not mean a product failed, but spares and support capacity should be contracted. Unmeasured media visibility does not mean no interest; it means this edition declines to compare that axis until the corpus review is complete.

Manufacturing Cell Checkpoint

A WRC visitor can compress the checkpoint into short booth questions. “How many did you make?” becomes “For which period, how many were produced, shipped for payment, and currently active?” “Do you have many customers?” becomes “Can you separate direct paid customers, pilot sites, and ecosystem partners?” Record the presentation, manual, filing, or follow-up URL and its date alongside every answer.

2.9 Limitations and update rules

First, the atlas is a snapshot of public evidence available through August 17, 2026. Private-company financing and customer contracts are selectively disclosed, while listed-company events after the reporting period may appear only in later filings. Second, official company material is useful for confirming product existence and support scope but may select favorable demonstrations and specifications. Third, resolving Chinese and English brands, affiliates, and listed vehicles is continuing work.

Fourth, the atlas refuses to collapse unlike product categories into one composite score. Readers do not receive an instant “top ten,” but they can choose criteria suited to their application. Fifth, WRC relevance can change until the event. Booth verification is a separate process requiring agreement between the newest official directory and floor plan. Inclusion among the 44 companies does not imply 44 confirmed booths.

Updates preserve historical definitions instead of overwriting them. A new financing round is checked for overlap with prior rounds. New revenue keeps its reporting period and consolidation scope. A shipment claim keeps its cumulative cutoff date. Media visibility must be recomputed with the same outlet set and story-eligibility rules for a newly defined window. An unresolved entity or value changes state only when the date and source of the resolving evidence are recorded.

Under these rules, the atlas becomes an auditable decision instrument rather than a scoreboard of impressive numbers. Readers can ask which source supports each claim and what can be compared before asking which company appears largest. That is the common grammar required to interpret technology layers and platforms in the following chapters.

What to Learn Next

Chapter 3 rearranges the company roster into components, robot bodies, perception, contact, data, simulation, models, and integration. The metric discipline from this chapter remains in force. Strong technical documentation does not guarantee high revenue; high shipments do not make a system suitable for every contact task. The next question is not “Who is bigger?” but “Which layer depends on which other layer, and where does a manufacturing-cell bottleneck arise?”

References

1. State Administration for Market Regulation (2026). National Enterprise Credit Information Publicity System. Chinese government registry.
2. GDELT (2026). GDELT DOC 2.0 API Documentation. Retrieval endpoint.
3. LimX Dynamics (2026). LimX official news archive. Official company website.
4. Geek+ (2026). Annual Results Announcement for the Year Ended December 31, 2025. Official issuer results.
5. Pudu Robotics (2025). PUDU Robotics Product Brochure. Official company brochure.
6. Deep Robotics (2025). DEEPRobotics Downloads. Official company website.
7. DexForce (2026). DexForce Documentation. Official company documentation.
8. Noitom Robotics (2026). Noitom Company Profile. Official company website.
9. PaXini (2026). PaXini and SynapX Strategic Partnership with OmniVTLA 2.0. Official partner website.
10. Reuters (2026). Reuters Site Search: Robot Company Corpus Retrieval. Publisher retrieval endpoint.
11. South China Morning Post (2026). Robotics Corpus Search. Publisher retrieval endpoint.
12. Xinhua (2026). Robotics Corpus Search. Publisher retrieval endpoint.
13. UBTECH Robotics (2026). 2025 Annual Report. Hong Kong Exchanges and Clearing filing.
14. Roborock (2025). Hong Kong Stock Exchange Application Proof. Hong Kong Exchanges and Clearing filing.
15. Ecovacs Robotics (2025). 2024 Annual Report. Exchange-hosted annual report.
16. MicroPort (2026). 2025 Annual Report. Official annual report.
17. Beijing Municipal Government (2026). ROPA6 Approval Notice. Government approval record.
18. World Robot Conference (2026). Official Exhibition Directory. Organizer website.
19. DexLink authors (2026). DexLink Hand: A Compact, Affordable, 16-DOF Linkage-Driven Hand with Human-Like Dexterity. arXiv:2606.17418.
20. ComSim authors (2026). ComSim: Building Scalable Real-World Robot Data Generation via Compositional Simulation. arXiv:2604.11386.
21. BestMan authors (2026). A Scalable Embodied Intelligence Platform for Seamless Real-to-Sim-to-Real Transfer of Household Mobile Manipulation Tasks. arXiv:2606.18646.
22. Physical Intelligence (2025). pi0.5: A Vision-Language-Action Model with Open-World Generalization. arXiv:2504.16054.
23. NVIDIA (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734.
24. DEXOP authors (2025). DEXOP: A Device for Robotic Transfer of Dexterous Human Manipulation. arXiv:2509.04441. Terry #10.
25. ByteDexter authors (2025). ByteDexter: Efficient Dexterous Manipulation Data Collection. arXiv:2507.03227.
26. ASAP authors (2025). ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. arXiv:2502.01143.
27. Imagine2Touch authors (2024). Imagine2Touch: Learning Contact-Rich Manipulation from Visuo-Tactile Imagination. arXiv:2405.01192.
28. Humanoid-Gym authors (2024). Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer. arXiv:2404.05695.
29. DexCap authors (2024). DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation. arXiv:2403.07788.
30. Zhengyou Zhang (2000). A Flexible New Technique for Camera Calibration. DOI:10.1109/34.888718.
31. Konstantinos Daniilidis (1999). Hand-Eye Calibration Using Dual Quaternions. DOI:10.1177/02783649922066213.
32. Radu Horaud et al. (1995). A New Technique for Fully Autonomous and Efficient 3D Robotics Hand/Eye Calibration. DOI:10.1109/70.476888.
33. Frank C. Park et al. (1994). Robot Sensor Calibration: Solving AX=XB on the Euclidean Group. DOI:10.1109/70.326576.
34. Benjamin W. Mooring et al. (1991). Robot Calibration. Wiley.
35. Roger Y. Tsai et al. (1989). Position and Orientation Calibration of a Robot Camera as a Pan-Tilt Unit. DOI:10.1109/70.34770.
36. Samad Hayati (1985). Geometric Approach for Robotic Calibration. DOI:10.1109/robot.1985.1087372.