H1 shipments rose by roughly 300% · but there is still no industry-wide dataset that shows actual working hours, human intervention and uptime together.
Actual Figure 02 operating hours directly confirmed by BMW2
?
How many humanoids are doing repetitive work at customer sites worldwide · no aggregate count exists
2026.09.15 · a new model arrived claiming 20 productive hours a day · Agility's Digit 5 · 90 minutes of operation with a 9-minute charge · “more than 20 hours of productive work in a 24-hour day” · but this is a company-stated performance claim and initial deliveries are scheduled for H1 2027.37 · hours already worked and hours a robot is said to be capable of working belong in different columns.
What this report counts · output robots actually produced while relying less on human hands.
※ Dates come in three layers · Published 2026.09.19 · Public-source review cutoff 2026.09.19 · The period covered varies by item (H1 shipments: Jan–Jun 2026 · robot density: 2024 · company disclosures: the date of each announcement) · the date shown with each item is the date relevant to that source.
SHIPMENTS ARE NOT WORKERSPAGE 02
THE FIRST OPTICAL ILLUSION
20,000 units are not 20,000 workers.
The top three boxes divide the same 100% by use case · the bottom two are evidence of field operation with different denominators.
Under Counterpoint's methodology, global H1 2026 shipments exceeded 22,000 units · up roughly 300% year over year1 · all figures below use this classification.
Estimates for the same half-year differ by almost twofold across institutions. A 2026.08.20 report from the China Humanoid Robot & Embodied Intelligence 100-Person Forum counted more than 40,000 units in China alone · equal to 97% of the global total43 · if it used the same definition as Counterpoint's 22,000-plus global figure released the same day, both figures could not be true at once · the difference implies different definitions of what counts as a humanoid, when a unit is counted as shipped, and whether small education or performance models are included · in an industry where even shipment counts diverge this sharply, “how many were sold” is not enough to judge commercialization.
Share of total shipments by use case
Total shipments
22,000+H1 2026 · up roughly 300% YoY
Outside manufacturing & logistics
79%+Research · performance · data 60%+ · service & guidance 19%
Manufacturing & logistics
18%
From here, the denominator changes · these are measures of field operation, not use-case shares
Repeat real-world work
?
Paid autonomous labor
?
※ 79%+ and 18% are not sequential filters · they are adjacent shares of the same 100% · the 18% is a use-case category—13% manufacturing plus 5% logistics—not a count of workers replaced.1 ※ Paid work can also occur outside manufacturing and logistics, so the 18% should not be read as the sole gateway to economically useful labor.
Why are the bottom two boxes blank?
79%+ is a lower bound: 60%+ plus 19% for service and guidance · the four use-case categories do not sum exactly to 100% because of rounding in the source · there is no global aggregate for repeat real-world work, only individual cases such as BMW, GXO, Longcheer and CATL · and we found no dataset for paid autonomous labor that discloses unit count, human-intervention frequency and uptime together · so the two boxes are intentionally left without a size.
Paid autonomous labor · customer budget is actually spent · real work is performed · and the robot operates without continuous human teleoperation.
There are statistics through the use-case layer · beyond that, the evidence becomes case-by-case.
Four reasons the numbers diverge
Category
Where definitions diverge
What to watch for
Form factor
Bipedal walkers only · or wheeled mobile humanoids as well
Do not compare shipment figures as if they were defined identically
Use case
Research · education · performance included in industrial sales
Separate product sales from labor substitution
Timing
Production · factory shipment · customer delivery · installation
Separate pre-installation inventory from production plans
Operation
Field trials · data collection · commercial operation
Being inside a factory does not by itself prove commercialization
The fact that no one knows the exact global count of working humanoids tells us where the industry stands today.In a mature industry, shipment volume is followed by uptime, maintenance cost and repeat orders · humanoids are still missing that second set of numbers.
THE COMMERCIALIZATION LADDERPAGE 03
WHEN DOES A ROBOT BECOME LABOR
When does a robot count as “working”?
A framework for reading stages · not a scorecard for companies
01
Video or stage demoControlled environment · can be edited
SHOW
02
PrototypeRepeated in company research facilities
LAB
03
Customer-site trialTested in someone else's factory · commonly called a PoC
TEST
04
Paid deliveryCustomer budget committed
PAID
05
Deployed in live operationsInstalled inside production or logistics workflows
NOW
06
Repeat real-world workContinuously handles real products or logistics
NOW
07
Long-duration autonomous workLow human intervention · high uptime
NEXT
08
Economics versus human laborReal hourly cost with all costs included
OPEN
09
Repeat orders in the hundreds to thousandsThe first customers expand their installed fleets
OPEN
10
Mass-production industryMany customers · standardization · service network
LATER
Most of the public evidence sits at Stages 5–6 · some tasks have entered real production flows · but to verify Stage 7, the public record still lacks human interventions per hour and mean time between failures (MTBF).
But these ten stages do not have to occur in sequence. Large supply contracts can arrive before customer economics are disclosed · and a small-volume specialized process can make economic sense without a mass order · as of 2026.05, Agility had secured more than $300 million in multi-year Digit 5 orders, while customer-level economics remained undisclosed37 · Early contracts in the hundreds to thousands of units have appeared, but the “repeat orders after operation” defined in Stage 09 are still unconfirmed · Stages 07 and 08 remain blank · contract size, operating maturity and economics are accumulating at different speeds.
AXIS A · WHO SAID IT
Nature of the evidence
Direct customer confirmation · company announcement · regulatory filing · local-government announcement · independent verification · who says it changes the weight of the evidence · A regulatory filing confirms that the filing was made and who made the claim; it does not mean the regulator verified the operating hours or future performance described inside it.
AXIS B · HOW FAR IT WENT
How far has it gone?
The ladder on the left is a framework for readers to interpret stages, not a ranking of companies · the first two stages are still outside the customer site, so real deployment starts at Stage 3 · a company can have strong sales evidence but weak labor evidence, or the reverse · reducing a company to one number hides these different dimensions of progress.
THE KEY DISTINCTION
Unitree · strong sales evidence / weak evidence on hours worked. Figure F02 · sales volume undisclosed / strong evidence of actual work. Agility · strong order evidence / customer economics undisclosed.
Even within one company, each generation must be verified separately · Figure 02 may have proven 1,250 hours at BMW, but Figure 03's long-run performance is a separate question · a new model does not inherit the operating record of its predecessor.2
THE EVIDENCE MATRIXPAGE 04
WHO HAS ACTUALLY WORKED
Who is actually ahead?
Judged by evidence of real work at customer sites—not by name recognition
AGILITY · DIGIT 4
65,000h+
The longest case in this comparison
The longest disclosed cumulative operating time belongs to Digit 4 · at GXO's Flowery Branch logistics center: more than 100,000 totes moved in aggregate · roughly 98% task accuracy while working37 · Digit 5, unveiled on 2026.09.15, is scheduled for initial deliveries in H1 2027.
Field operation confirmed · customer economics undisclosed3
FIGURE · F02
90,000+
Actual BMW parts handled
1,250 hours · supported production of more than 30,000 BMW X3 vehicles · directly confirmed by customer BMW · the BMW line's stated requirements were an 84-second full cycle, 37-second load time and 5 mm tolerance.26
Directly confirmed by customer · contract terms undisclosed2
AGIBOT · G2
8 units · 64h
No. 1 in shipments, but the labor evidence covers eight units
The company ranked No. 1 globally with 9,700 H1 shipments1 · the disclosed labor evidence is eight robots deployed on Longcheer's tablet line · 2026.06.23–28 · six days · 64 hours of line operation · humans handled inspection, adjustment and exceptions.
Company-by-company public-evidence comparison · reviewed through 2026.09.19
Company
Field operation
Commercial transaction
Customer economics
Nature of the evidence
Agility
Digit 4 · 65,000+ cumulative operating hours3 · GXO Flowery Branch · 100,000+ totes moved · roughly 98% task accuracy while working37
Paid Robot-as-a-Service (RaaS) contracts · as of 2026.05, more than $300m in multi-year Digit 5 orders · anchored by an undisclosed customer's three-year, 1,000-robot agreement · subject to contract milestones3738 · the company itself recorded $1.78m in 2025 net revenue38
The most-shipped robot is still not the robot with the most documented work. The four columns are not collapsed into one grade · time spent operating in the field · whether money changed hands · whether the customer gained economically · and who made the claim are different questions. The conclusion of the table is that the third column is blank in all nine rows. Field operation and commercial transactions come with names and numbers · not one case discloses how much the customer actually saved.
Company notes · the traps of model generations and contracts
BMW's stated targets for Figure were a success rate above 99% per shift and zero human intervention · but both are requirements, not disclosed achieved results · requirements should not be read as outcomes26 · even the operating period differs: BMW says 10 months, Figure says 11226 · Figure's strong 1,250-hour evidence applies to F02 · F03 still has to establish its own long-run operating record · The same generation-by-generation distinction applies to Agility · the 65,000 hours and roughly 98% GXO task accuracy are records for Digit 4, not for Digit 5 unveiled in 2026.09 · 98% accuracy should not be reinterpreted as 98% uptime or a 2% human-intervention rate37 · Agility's 65,000 hours are not broken out to show how much was paid, autonomous production time · Tesla has few disclosed external-work performance metrics relative to the scale of its plans · Unitree has commercialized products, but evidence of commercialized labor is a separate matter.
An industry where some of the most concrete labor evidence comes from less famous companies.
WHAT THE ROBOTS ACTUALLY DOPAGE 05
A FEW JOBS BEFORE A GENERAL WORKER
What are humanoids actually doing today?
A handful of dull, structured jobs arrived before the general-purpose worker
Concept image · not footage from a specific site
The jobs with verified spending and operating hours have common traits · locations and objects are relatively fixed · repetition is high · and the work is physically taxing or monotonous for humans.
01
Picking parts · loading jigs
BMW · Figure · picks sheet-metal parts and places them into welding fixtures · the strongest customer-confirmed case.
02
Moving boxes and materials
GXO · Digit · repeatedly moves standardized totes between conveyors and racks.
03
Sorting and transfer
Longcheer · AgiBot · moves and sorts units in a tablet inspection process · six days and 64 hours of continuous line operation.
04
Sequencing parts for workers
Atlas, Apollo and others · deliver required parts to workers in sequence · still largely in field-trial mode.
05
Precision assembly and wiring
Flexible cables · tool changes · tiny screws · a wall of tactile sensing and exception handling.
06
Home cleaning
Tau Robotics · paid humanoid cleaning in San Francisco · $30 per hour · but a trained operator at another location teleoperates the robot in real time · video from visits is used for model training30
Paid service ≠ autonomous labor. The $30 is the customer's service price, not the cost of one autonomous robot-hour · the company's terms state that recorded video and operator commands become training data for its vision-language-action models30 · the cleaning service has been sold; autonomy has not.
THE ACTUAL ENTRY POINT
Not a robot that does everything a person can · a robot taking on a few tasks inside spaces designed for people.
For the humanoid form factor to have economic meaning, it must be able to adapt to more tasks while requiring fewer changes to the environment than existing AMRs, conveyors and robot arms · Humanoids do not compete only with people · they also compete with specialized robots already doing the work.
Structured · repetitive welding
Specialized robots already do it
HD Hyundai Heavy Industries · a collaborative robot built on an electric tracked autonomous mobile platform · used in actual production · throughput rises from roughly 500 tons per day with manual work to about 750 tons in daytime operation and up to 1,000 tons with day-and-night operation · one person supervises multiple units on a process previously handled by seven skilled workers.32
DIDEN Spider 30 uses magnetic feet to traverse vertical walls and curved surfaces as a quadruped33 · Persona uses a bipedal platform designed to move stably through a shipyard34 · a 2026.07 welding demonstration is confirmed · but the company explicitly described it as teleoperation in its own disclosure35 · the task is the same—welding—but the chosen body form differs.
The reason to look human may not be to imitate humans · it may be to enter human-built environments with fewer modifications.
THE BRAIN GOT BETTER · THE BODY STILL BREAKSPAGE 06
WHY NOW · WHY NOT YET
The way we teach changed before the body did.
A brain that sees · understands · acts · but a factory demands 100,000 repetitions, not one successful demo
Traditional industrial robots were closer to perfectly memorized choreography · newer robots increasingly see scenes through cameras · understand language instructions · and translate them into joint movements · this family of models is called vision-language-action models (VLA).
SEE
SEEPerceive objects · people · space through cameras and sensors
UNDERSTAND
UNDERSTANDTurn language and context into a single task objective
ACT
ACTGenerate continuous arm · hand · leg movements
LEARN AGAIN
LEARNLearn again from simulation · teleoperated demonstrations · real factory failures
The acceleration of the past two or three years did not come from one invention · imitation learning · simulation · GPUs · EV supply chains · and China's precision-component manufacturing capacity all moved at once.
Why is data the bottleneck? · Jensen Huang said in a 2026.06 keynote that “the biggest problem is data”, arguing that most of the world's video is third-person while robot AI needs first-person data from the robot's point of view · his explanation was that the gap must initially be filled with teleoperated demonstrations, simulation and synthetic data · “in physical AI, compute becomes data”45.
※ NVIDIA is a direct participant in this market, which should be kept in mind.
That is why money is now flowing not only into models, but into human behavior · on 2026.08.25 Figure launched Index, a collection program in which people record and upload their own actions using smartphones31 · on 2026.09.03 it announced a partnership with Nscale covering up to 100,000 GPUs42 · the cost of collecting data and the cost of training on it are rising together.
Does pouring in money and data automatically produce dexterity? · the counterargument. Rodney Brooks argues that human dexterity depends heavily on touch, information that video data does not contain · “collecting only visual data is not collecting the right data”44 · the published work records seen earlier are the numbers that test this objection · narrowly defined work has already begun · public evidence of general-purpose dexterity has not.
Published figures for Index and Nscale
Figure Index · company-reported figures · 16m+ uploaded human-action videos · 44,000+ weekly active contributors · 100+ countries · $15m in cumulative contributor payouts31 · the dataset spans cooking, cleaning and laundry at home as well as real activity in warehouses, restaurants, factories and offices · human behavior → physical-world data → model → deployed robot → failures and exceptions → back to data.
Nscale partnership42 · up to 100,000 GPUs · $3.5bn initial compute commitment · first deployment targeted for H2 2027 · Nscale becomes a Figure shareholder · this does not mean 100,000 GPUs are running today or that Figure raised $3.5bn in cash; it is a compute commitment · the manufacturer's model-compute spending and a customer's cost of adopting robots belong to different accounts, so this amount is not added to the customer-side hourly cost on the next page.
Training teleoperation and demonstrations ≠ human intervention during actual work · a person moving to teach a robot is a training method · a person rescuing a failed robot at a customer site is operational intervention and hidden labor cost · Index videos are human-action data, not robot operating hours or teleoperation hours.
Picking up a cup once and picking up the same object 100,000 times in a factory are different problems.
01 · DEXTERITY
Hands and touch
The robot must sense slip, material and force differences in real time · a core bottleneck in handling human tools and deformable objects.
02 · INTERVENTION
Human rescue
If a remote operator has to step in after a failure, labor cost returns · a teleoperated demonstration for training and an intervention during live operation are different, yet disclosures often fail to separate them27 · we found no public data on the number of human interventions per hour.
03 · MTBF · mean time between failures
Keep running without failure
Joints · reducers · cables · sensors must repeatedly withstand shock and heat · a core metric for industrial machinery.
04 · CYCLE TIME
Move as fast as a person
Being able to do a task is different from matching human production speed · if hourly throughput is low, even a cheap robot becomes expensive labor.
05 · ENERGY
Uptime matters more than charge duration
Two hours versus four hours matters less than real production uptime, including automatic charging and shift operation.
06 · INTEGRATION
The factory must change too
BMW also modified safety zones, communications and IT systems · the investment is in the entire workflow, not just one robot.
Even at a 99.9% success rate · 100,000 cycles imply roughly 100 failures in theory.The key is not the number of failures but how they are recovered · if the robot recovers on its own, that is autonomy · if a person has to intervene, that is hidden labor cost.
THE PRICE OF ONE PRODUCTIVE HOURPAGE 07
ROBOT HOUR COST
Is a cheap robot really cheap?
Not the purchase price · the cost of one hour needed to produce the same output, including the human share
The cost of one robot hour · EFFECTIVE PRODUCTIVE ROBOT HOUR COST
All costs attributable to the robot over a period ÷ effective operating hours normalized to the same output as a human worker
NUMERATOR · acquisition cost allocated to the period (depreciation) · maintenance · parts · batteries · energy · software · labor for supervision · recovery · teleoperation · insurance · safety equipment · factory modifications DENOMINATOR · scheduled operating hours × actual uptime × throughput relative to a human × good-output rate GOOD-OUTPUT RATE · relative value with human work set to 100 RESULT · dollars ÷ hours = effective hourly cost · directly comparable with hourly human labor cost Why multiply four terms? · downtime · slower speed · defects all reduce the “hours actually usable to produce the same output.”
Autonomy is a separate metric · intervention-free productive hours ÷ scheduled production hours · output is not discounted simply because a human is present · if the robot makes 100 units under supervision, all 100 units count as output · the supervisor's labor goes into cost · the economics can still work if one person supervises ten robots.
Illustrative Digit Robot-as-a-Service (RaaS) assumptions included by Agility in a U.S. SEC filing · $8,500 per month · 120 hours per week · about 520 hours per month(120 hours × 4.33 weeks) gives roughly $16.35 per hour · but this is a company-provided illustration, not an actual customer ledger.4
Minimum calculation using lease fees only · this includes only the monthly lease, not the “all costs” in the formula above
Uptime
Relative speed
Effective cost
100%
100%
$16.35
80%
100%
$20.43
80%
70%
$29.19
60%
70%
$38.92
50%
50%
$65.38
Formula · 16.35 ÷ (uptime × speed relative to human) · the 100% in the first row does not mean 24/7 operation; it means all contracted 120 hours per week are used · 120 hours is already about 71% of the 168 hours in a week, so the assumption already implies near-continuous use. Against the roughly $30.50/hour U.S. material-moving wage benchmark Agility used in the same filing · uptime × relative human speed must exceed roughly 54% for the robot to come out ahead on this simplified cost basis.4This table is a lower bound using only the lease fee · the good-output rate is assumed to be 100% · if customer-side supervision, integration or safety-equipment costs are not included in the RaaS fee, the actual cost rises further.
Keeping a robot from stopping can matter as much as cutting its price in half.
What BMW and Figure disclose is 90,000+ parts · 1,250+ operating hours · support for 30,000+ X3 vehicles and that the task was loading sheet-metal parts2 · to fill the denominator in the formula above, we would also need a human worker's throughput on the same task and how many times humans intervened · neither is disclosed, so this case does not allow an effective hourly cost calculation.
United States · two benchmarks
$30.5 vs $48.6
$30.50 is the material-moving labor assumption Agility selected for its own economics illustration4 · $48.62 is the June 2026 U.S. private-sector manufacturing employer cost for total compensation (wages $32.50 + benefits $16.12)13.
Series · the first is a company-selected comparison assumption · the second is a broad statistic on actual employer compensation costs · the economic threshold changes depending on which benchmark is used.
Korea · total manufacturing labor cost
Lower than the U.S.
A monthly total labor-cost statistic including all employer-borne costs · economics may still work in large-scale manufacturing and shift-based processes.
Series · monthly data · converting it to hourly terms requires an assumption for monthly working hours, so no converted figure is shown.14
China · production-worker wages
Lowest of the three
Average wage for production and manufacturing workers · annual series · a high-priced U.S.-style robot lease faces a much harder comparison if carried over unchanged.
Series · wage data excluding social contributions · conversion would require assumptions for exchange rates, annual working hours and employer contribution rates, so no hourly figure is shown.15
※ The three figures come from different statistical series and cannot be placed side by side as hourly equivalents · only the U.S. figure remains numeric because Agility itself used it as an hourly benchmark · for Korea and China, only the relative ordering is shown · stating that the series differ does not make them directly comparable.
※ Economics does not end with an hourly-wage comparison · in processes constrained by shortages of skilled labor, “does it unlock production capacity when people are unavailable?” matters to the investment case alongside “is it cheaper than a person?” ※ We have not found a public case where a humanoid has quantified this additional value all the way through to an actual customer return on investment.
We found no public case that proves humanoid economics versus human labor with actual on-site numbers. The $16.35 figure above is an illustrative assumption included by Agility in a filing, not a realized hourly cost taken from a customer's books. Autonomy and economics belong in different ledgers. Economics can work even with limited autonomy if supervision labor is included and the numbers still add up · conversely, an hour with zero intervention can still be uneconomic if throughput is too low · the specialized robot on the previous page also works by having one person supervise multiple machines on a process once performed by seven skilled workers32 · the same yardstick should be applied to humanoids · humanoids do not have to eliminate humans completely to be economically viable.
CLOSING THE LOOPPAGE 08
THE UNITED STATES · CHINA · KOREA
Who closes the loop first?
Field data → reliability → price → repeat orders · the U.S. and China are trying to close the loop from different starting points
Reducing U.S.–China competition to “America does AI · China does hardware” misses the point · the real contest is over this loop: field data → reliability → price → repeat orders—specifically, who can close it first. The two countries have each captured different links first.
THE UNITED STATES
Starting with field data
First link captured · field data and reliability · operating hours are accumulating at named customer sites such as BMW, GXO, Mercedes and Jabil
What supports it · vision-language-action models · general-purpose foundation models · simulation
Link not yet closed · price · high hardware cost and low production volume
Key question · can a robot that works well be manufactured at the economics of tens of thousands of units?
CHINA
Starting with price
First link captured · price and mass production · motors, batteries and precision manufacturing from EV and smartphone supply chains carry over directly
What supports it · all five leading H1 2026 shippers were Chinese companies · combined share 86%1 · in 2026.09 UBTECH began operating a dedicated Liuzhou factory with planned annual capacity of 10,000 units40
Link not yet closed · field data · limited disclosure of labor-performance metrics leaves the first link incomplete
Key question · can tens of thousands of robots become the equivalent of tens of thousands of workers?
KOREA · THE MIDDLE WITH A FACTORY
Korea has factories that could fill the first link and mass-production experience that could support the third link · but its share in complete humanoids is small and self-sufficiency in core components is weak · the next page examines which link Korea can realistically capture.
※ The three country profiles above describe conditions on the robot-maker side · if a Chinese robot enters a U.S. factory, the economic threshold is set by U.S. labor costs · the manufacturer's nationality and the customer factory's nationality should not be treated as the same thing. ※ The Liuzhou factory is evidence of manufacturing capacity · it does not mean 10,000 annual sales or 10,000 autonomously operating units at customer sites.
Concept image · not footage from a specific site
KOREA · THE ROBOT COUNTRY WITHOUT ALL THE PARTSPAGE 09
A STRANGE STARTING POINT
Is Korea late?
The world's densest user of robots in manufacturing · yet less than half of core humanoid components are domestically sourced
1,220units
industrial robots per 10,000 manufacturing workers · No. 1 globally in 2024 · more than nine times the world average of 132 · U.S. 307 · China 166 · EU-27 231.16
45%
domestic sourcing rate for core components stated by the government on 2026.08.27 current level · 80% target by 2030 · actuators, hands and sensors identified as weak links · the source does not specify the base year for the current figure.17
Korea is not a country introducing robots to manufacturing for the first time · the question is whose brains and components will power the next generation.
Concept image · not footage from a specific site
Korea's biggest asset may be less its robot brands than its real factories · physical-world data such as skilled workers' motion trajectories, force control, failures and exception handling still cannot be fully reconstructed from web text and video alone · but crowd-sourced video collection has also emerged, so simply owning factories does not create an automatic advantage · the edge exists only if real workplaces are captured in a form that can train robots.
THE FACTORY DATA EDGE
Even without building the largest AI model · Korea may be able to accumulate high-quality data on “how work is done.”
Being No. 1 in robot density does not mean Korea is only now starting automation · it means the next layer is being built on top of an installed base · annual new installations, however, have not increased materially since 201916 · conventional automation is mature · humanoids are being tested as a potential new layer on top of it.
2026.07 · turning shipyard welders' know-how into data begins · a shipbuilding AX consortium led by Rainbow Robotics, with Pusan National University's School of Advanced Convergence, Sungwoo Hitech, OpenPass Robotics and Busan IT Industry Promotion Agency · customer HJ Shipbuilding & Construction provides a real test environment plus skilled welders' motion and design data · multiple cameras and sensors record the work for conversion into 3D training data.36
※ The 2026 goal is to build the data pipeline and test the feasibility of shape recognition and action generation · this does not mean autonomous welding has already been commercialized.
Automotive-style mass production · quality control · safety · service networks also become more valuable as humanoids move from prototypes to industrial goods · Korea is both a demand market and a manufacturing base · but also a country with weak self-sufficiency in components.
01
Complete robotsBoston Dynamics · Rainbow · ROBOTIS · Korean startups
Early customersHyundai Motor · Kia · electronics · batteries · shipbuilding · logistics
04
Factory dataFailures · exceptions · work hours · skilled-worker trajectories
05
Back into mass productionData feeds back into the next generation of robots and components
KOREA · WHO GETS PAID FIRSTPAGE 10
EVIDENCE BEFORE THEME
Who reaches revenue first?
Classified not as “robot-related stocks,” but by the strength of real customers, supply contracts and field trials · this is about how directly industry growth connects to revenue, not the attractiveness of any stock.
Connected · DIRECT CONNECTION
Customer demand and supply are already connected
A real customer for the complete robot or a disclosed component supply line already exists · this tier has the shortest bridge between industry growth and revenue.
Hyundai Mobis · Atlas actuator collaboration18 SBB Tech · KRW 1.92341bn precision-reducer supply contract for humanoid robots · 26.7% of recent annual revenue19 · A multi-year contract running from 2026.08 to 2028.04 · it should not be read as one year of revenue.
In field validation · FIELD TEST
The product has reached a real operating site
The technology has entered factories and logistics centers, but repeat orders and long-duration operating-hour evidence are still limited.
ROBOTIS · The actuator business already generating sales and the AI Worker platform undergoing field validation are at different stages · actuators are already sold · the platform is undergoing field trials at BGF Retail21and CJ Logistics28.20 Rainbow Robotics · brought under Samsung22 · MOU with CJ Logistics to co-develop humanoid logistics robots46
Option value · STRATEGIC OPTION
Strong assets, but the revenue link is a separate question
A long-term option combining AI · appliances · service robots · data infrastructure · separate from current industrial-labor performance metrics.
The technologies—reducers, motors and sensors—exist, but public sources do not confirm large humanoid-specific orders or customer links · the names below are candidates based on technical capability, not confirmed suppliers.
Korean precision-component makers producing reducers · motors · sensors · apart from Hyundai Mobis and SBB Tech in the “Connected” tier above, we found no other publicly confirmed humanoid-specific supply connection.
THE RULE
“Can make it” and “is shipping it” are different · the size of an industry and the degree to which it feeds a company's revenue are different too.
Major Korean field trials · disclosed progress differs by site · latest public status shown for each row
Industry
Customer · robot company
Task
Current stage
Automotive
Hyundai Motor · Boston Dynamics
Parts sequencing → future assembly
Training and validation · target deployment at North American plants in 202811
Electronics
Samsung Display · Rainbow Robotics
Lens and jig replacement · opening and closing equipment doors
Field trial · status as of 2025.10 · no later stage found in public sources25
Shipbuilding
HD Hyundai Heavy Industries · Airobot
Fire monitoring in ship bulkhead sections
Government demonstration project in progress · status as of 2026.0521
Shipbuilding
HD Hyundai · Persona AI
Shipyard-specific welding
Joint development and validation under way · bipedal platform · prototype due by end-2026 · field trials planned for 2027 · autonomous robot operating hours undisclosed · 2026.0334
Shipbuilding
Samsung Heavy Industries
Welding · hot-work monitoring · autonomous patrol
Field trial · status as of 2025.10 · no later stage found in public sources25
Refining
SK Energy · Holiday Robotics
Valve operation in toxic environments
Government demonstration project in progress · status as of 2026.0521
Logistics
BGF Retail · ROBOTIS
Product re-sorting inside a logistics center
Government demonstration project in progress · status as of 2026.0521
Logistics
CJ Logistics · ROBOTIS
Replenishing cushioning material on the packing line at the Gunpo fulfillment center
Field validation since 2025.09 · expansion plan announced for 202628
Appliances · electronics
LG Electronics · CLOiD
Deployment into production processes
Planned for H1 2026 · actual deployment not confirmed · status from 2026.04 earnings call29
Korea has many factories and more trials, but the publicly disclosed pace of progress differs widely. Hyundai Motor targets deployment in 2028 · HD Hyundai Heavy Industries, SK Energy and BGF Retail had government-backed demonstration projects confirmed through H1 2026 · HD Hyundai and Persona AI have started joint-development validation · CJ Logistics has outlined expansion after field trials · LG Electronics remains at the deployment-planning stage · and no next stage after the 2025 trials at Samsung Display and Samsung Heavy Industries was found in public sources · even within one country at the same time, some sites disclose their next step while others do not.
How to read the table · government plans and the meaning of “not confirmed”
Government plan dated 2026.08.27 · build mass-production systems for industry-specific humanoids across ten sectors by 2030 · make demand sites the focal point of development · government purchases of 250 domestically produced humanoids in 2027 and 1,080 cumulatively by 203017 · however, the document does not list individual corporate customers · the stated purchase and distribution targets are research institutions such as engineering universities and government-funded research institutes · so the numbers do not imply equivalent deployment in commercial factories · the same distinction between research shipments and actual labor also applies to government purchases.
“Not confirmed” in the table means no supporting public source was found; it does not mean the company stopped the project.
In the table above, the tier closest to actual revenue is components, not complete humanoids · that is the only tier with a disclosed supply contract · field trials are increasing, but no row yet shows repeat orders or long-duration operating hours.
Government plans, budget proposals and clues on employment effects
2026.09.10 · component budget · the government proposed KRW 62.0bn in the 2027 budget for sensors · actuators ·secondary batteries for humanoids · more than five times the KRW 11.6bn budget for 202641 · this is a budget proposal, not disbursed funding · and it covers those three component categories, not the entire robotics budget.
Employment effects · in Korea, the first impact may be less about mass layoffs and more about keeping hard-to-staff · dangerous · repetitive operations running · some frontline roles may shift from performing tasks to robot training, supervision, maintenance and exception handling · but this is a hypothesis about Korea's early adoption path, not a general law · this report does not measure employment effects themselves, and neither “robot adoption immediately causes mass layoffs” nor “robot adoption immediately causes a productivity boom” is established by the evidence reviewed here.
Factory-data advantage · a potential advantage, not a proven one · public sources do not show how much long-duration operating data has actually accumulated.
Two paths diverge · U.S. AI + Korean manufacturing data + Korean components and mass production. Or · adoption centered on foreign complete robots while Korea remains mainly a customer market and assembly base.
THE INVESTOR'S MAPPAGE 11
WORKED HOURS BEFORE A STORY
What should we watch next?
Not a buy recommendation · metrics that can verify the direction of the industry
6~18M
Near-term clock · the threshold from pilots to financial results
Field trials become paid contracts · does a customer budget support the test?
Repeat orders · do five units become 50, then 500?
Robot operating hours · actual hours worked in real factories, not production plans
Component orders · do prototype orders turn into mass-production awards?
Humanoid revenue · on the income statement, not just in presentation slides
3~10Y
Long-term clock · conditions for a winner
Field data · a loop that accumulates real failures and recoveries
Autonomy · how much remote human labor can be stripped away
Reliability · mean time between failures · uptime · service network
Mass production · ability to build 10,000 units at consistent quality
Factory integration · integration into manufacturing and warehouse-management systems, safety and maintenance
PATH A · signals of deeper industrial adoption
Real customers and real operating hours appear
Customer sites such as BMW, GXO, Longcheer and CATL · lower cost of teaching new tasks through vision-language-action models and imitation learning · price declines from Chinese supply chains · labor shortages in advanced economies · simultaneous progress across technology, cost and demand.
PATH B · signals the industry remains stuck in pilots
Capable—but perhaps without a compelling reason to use it
For a single structured task, conventional robot arms, AMRs or conveyors may be cheaper and faster · industrial robot arms already cost in the tens of millions of won and have proven reliability and throughput · risks include persistent teleoperation dependence · failure to improve MTBF · excessive factory-modification cost · and field trials that never convert into repeat orders.
Only when these five metrics begin to be disclosed can Stages 7–9 on the ladder be verified in practice. 07 · autonomous hours · interventions · failures / 08 · realized hourly cost / 09 · repeat orders
01
Intervention-free productive hoursAutonomy metric · time spent handling real products without human interventionProductive Autonomous Hours
02
Total human interventionFrequency of rescue plus time per intervention · support staff required per fleet of robotsIntervention Rate & Support Ratio
03
Time running without failureUninterrupted runtime and uptimeMTBF · Uptime
04
Repeat ordersThe first customer's second orderRepeat Order
05
Realized hourly costEconomics metric · effective hourly cost including supervision and recoveryRealized Robot Hour Cost
Next near-term checkpoint · Agility has scheduled an Analyst & Investor Day for 2026.10.0638 · a chance to revisit milestone conditions on the $300m+ order book · production-scaling plans · and the business model · the content of that presentation is still in the future.
Shipments are the sixth metric.It is already the fastest-growing number · what matters next is the share of shipped robots that turn into actual labor.
WHAT THE NEXT DATA WILL ANSWERPAGE 12
EVIDENCE AND WHAT CHANGES IT
What will the next evidence answer?
Public evidence sits halfway up the ladder · the report closes with what would change that assessment
How far has it come?
Stages 5–6
Humanoids have entered real workflows for some repetitive manufacturing and logistics tasks · public validation of general-purpose work remains limited.
How fast is it growing?
H1 shipments +300%
22,000+ units · roughly 300% year-over-year growth · and more cases where repetitive work can be verified at customer factories.1
What is still missing?
Stages 7 and 8 are blank · contracts arrived before Stage 9
We found no public case proving long-duration autonomous operation or realized customer economics · large orders have already appeared, but repeat orders after operation—and delivery and operation of those volumes remain unconfirmed.
What would change the assessment?
Any one of these disclosures could change the Stage 5–6 assessment
A customer discloses autonomous operating hours as a number · the first customer expands its fleet · contract milestones turn into verified deliveries.
They have started working · but evidence of sales · work performed · and economics is accumulating at different speeds.
Unitree's Shanghai listing on 2026.08.197 · Agility's public-listing process in 2026.0938are bringing previously private sales and financial data into public markets · the industry is entering a phase with more numbers that companies are required to file, rather than only metrics companies choose to disclose.
※ This report does not judge whether current prices are fair · sector market capitalization and share-price gains24alone cannot tell us what level of autonomy, profit or adoption the market is pricing in · that judgment requires peer sets, the value of existing businesses, robot-related revenue and earnings assumptions—and belongs in a separate valuation study.
THE OPEN QUESTION
The more important question is not how many units shipped · but how much those robots produced while relying less on human hands. The question that remains · is this productivity enough to justify the contracts and investment now being committed?
Source ledger · evidence behind the figures in the report
1 · Counterpoint Research · H1 2026 global shipments of 22,000+ · roughly +300% YoY · use-case classification · shipment share by company · Original
2 · BMW Group · Figure 02 operating hours · parts handled · support for X3 production · Original
5 · Longcheer Technology · 8 AgiBot Lingxi G2 units deployed · 2026.06.23–28 · 6 days · 64 hours of continuous production-line operation · Original
6 · CATL · Galbot S1 in regular operation on a battery production line · Original
7 · Shanghai Stock Exchange · “Announcement on the Listing and Trading of Unitree Technology Shares on the STAR Market” 2026.08.18 · states that Unitree common shares would begin trading on 2026.08.19 · ticker 688836 · Notice
8 · Mercedes-Benz · Apptronik Apollo factory trial · Original
9 · Jabil · Apollo manufacturing and field-validation collaboration · Original
10 · Boston Dynamics · Atlas productization · customer adoption commitments · Original
11 · Hyundai Motor Group · Atlas manufacturing-site roadmap · 2028 parts-supply plan · Original
12 · Tesla / SEC · Optimus production-line and data-collection preparations · Original
13 · U.S. BLS · Employer Costs for Employee Compensation (ECEC), June 2026 · private-sector manufacturing $48.62 per hour (wages $32.50 + benefits $16.12) · all private industry $46.89 · Original
14 · Ministry of Employment and Labor · Korea Enterprise Labor Cost Survey · Original
15 · National Bureau of Statistics of China · wages for production and manufacturing workers · Original
16 · International Federation of Robotics · released 2026.04.08 · 2024 robot density per 10,000 manufacturing workers · world average 132 · Korea 1,220 · Singapore 818 · Germany 449 · Japan 446 · U.S. 307 · China 166 · EU-27 231 · China installed 295,000 new robots in 2024, 54% of the global total · operational stock roughly 2 million · Original
17 · Korea Policy Briefing · humanoid ecosystem and domestic-component strategy · announced 2026.08.27 · current level 45% → target 80% by 2030 · mass-production systems for sector-specific humanoids across ten industries by 2030 · government purchases of 250 units in 2027 and 1,080 cumulatively by 2030 · Original
18 · Hyundai Mobis · strategic collaboration on Atlas actuators · Original
19 · SBB Tech · Financial Supervisory Service electronic disclosure system (DART) · single sales/supply contract signed 2026.08.24 · KRW 1.92341bn of precision reducers for humanoid robots · Korean robot manufacturer (name undisclosed) · contract period 2026.08.21–2028.04.29 · 26.7% of recent annual revenue · Filing
21 · Ministry of Trade, Industry and Energy · “One-Year Achievements of the AI Robot Subcommittee” 2026.05.29 · specifies task, robot company and customer for each demonstration project · fire monitoring in ship bulkhead sections (Airobot · HD Hyundai Heavy Industries) · welding in curved ship sections (DIDEN Robotics · HD Hyundai Samho) · valve operation in toxic environments (Holiday Robotics · SK Energy) · product re-sorting in logistics centers (ROBOTIS · BGF Retail) · Original
22 · Samsung Electronics · Rainbow Robotics becoming a consolidated largest-shareholder affiliate · Future Robotics Office · Original
23 · LG Electronics · CLOiD · robot actuators · Bear Robotics · Original
24 · iM Securities · “Robotics Q2 Review” 2026.08.19 · Lee Sang-soo · Korean robotics-sector market capitalization (as of 2026.08.18) · August sector performance · brokerage research report
25 · ETNews · “Seize the Golden Time for Manufacturing AX <4> Humanoids Move Into Manufacturing Lines · Six Field-Demonstration Units in Operation” 2025.10.26 · companies, tasks and stages across six Korean field trials · Article
26 · Figure AI · 「F.02 Contributed to the Production of 30,000 Cars at BMW」 · requirements full cycle 84 seconds · loading 37 seconds · tolerance 5 mm · >99% success rate per shift · zero human intervention · ACHIEVED VALUES 1,250+ operating hours · 90,000+ parts · 30,000+ X3 vehicles · 11 months · Original
27 · Agility Robotics · “Digit Moves Over 100,000 Totes in Commercial Deployment” 2025.11.20 · 100,000+ totes at GXO Flowery Branch · intervention rate, uptime and operating mode undisclosed · teleoperated demonstrations mentioned only as part of policy learning · Original
28 · CJ Logistics Newsroom · “Industry-First AI Humanoid Robot Field Trial · MOU with ROBOTIS” 2025.09.25 · deployed from mid-September 2025 on the packing line at the Gunpo fulfillment center · replenishing cushioning material · plan to expand to major logistics centers from 2026 after year-end validation · Original
29 · LG Electronics Q1 2026 earnings conference call 2026.04.30 · “We plan to deploy CLOiD into PoC validation work starting in the first half of this year and are preparing production for that purpose” · home-robot market entry targeted for 2028 · Article
30 · Tau Robotics · website and service privacy terms · flat $30/hour humanoid cleaning service in San Francisco · invitation-only with a waitlist · “a trained person at another location controls the robot in real time and views the home through the robot camera” · “recorded video, telemetry and control commands become training data for the vision-language-action models we build · this is not a side effect but a core purpose” · Original
31 · Figure AI · “Introducing Index” 2026.08.25 · company-reported figures · 264,000 app downloads · 100+ countries · 44,000+ weekly active contributors · 16m+ uploaded videos · $15m in cumulative contributor payouts · footage shot by people on smartphones showing human actions · not robot operating hours · Original
32 · AsiaToday · “Robots Weld, AI Decides · The Future of the AI Shipyard Seen at HD Hyundai” 2026.06.14 · collaborative robot built on an electric tracked autonomous mobile platform · manual work roughly 500 tons/day → about 750 tons in daytime operation → up to 1,000 tons with day-and-night operation · one person supervises multiple units on a process previously handled by seven skilled workers · a collaborative robot, not a humanoid · Article
33 · DIDEN Robotics · company product description · “a quadruped work robot with magnetic feet · traverses stairs and obstacles inside ship blocks and moves across vertical walls and curved steel structures” · expansion into welding and inspection is described as a future capability, not a claim of actual production deployment · not a humanoid · Original
34 · Smart Maritime Network · “HD Hyundai begins verification phase for shipyard welding robots” 2026.03.23 · joint-development agreement among HD Korea Shipbuilding & Offshore Engineering, HD Hyundai Robotics, HD Hyundai Global R&D Center and Persona AI · “a bipedal humanoid platform capable of moving stably through a shipyard” · progression from the 2025.05 MOU · prototype by end-2026 · field trials and deployment planned for 2027 · Article
35 · Persona AI · company disclosure dated 2026.07.10 · Gen 1 humanoid welding demonstration at ARC Specialties in Houston · “the welding was performed through expert teleoperation, with an engineer controlling the robot in real time through a VR headset” · “the robot performed the motion, but all judgment remained human” · not a record of autonomous welding · Article
36 · InTheNews · “Pusan National University Consortium Selected for Shipbuilding AX Project · Digitizing Skilled Welding” 2026.07.29 · led by Rainbow Robotics · participants include Pusan National University's School of Advanced Convergence, Sungwoo Hitech, OpenPass Robotics and Busan IT Industry Promotion Agency · customer HJ Shipbuilding & Construction provides a validation environment plus skilled welders' motion and design data · recorded with multiple cameras and sensors and converted into 3D data · 2026 goal is to build the data pipeline and test feasibility of shape recognition and action generation · Article
37 · Agility Robotics · official “Digit 5” announcement 2026.09.15 · payload 22.7 kg · 90 minutes operation · 9-minute charge · 10:1 operating-to-charging ratio · “more than 20 hours of productive work in a 24-hour day” is a company-stated performance claim · initial deliveries H1 2027 · general North American availability by end-2027 · the 65,000+ cumulative hours, 100,000+ totes at GXO Flowery Branch and roughly 98% task accuracy while working are all Digit 4 records · as of 2026.05, Digit 5 had more than $300m in multi-year orders subject to contract milestones · safety standard ISO 25785-1 · ANSI/A3 TR R15.108 is under development · Original
38 · Agility Robotics · Churchill Capital Corp XI · SPAC merger · Form S-4 registration statement filed with the SEC on 2026.09.04 · $2.5bn enterprise value · 2025 net revenue $1.78m · the $300m+ order book is anchored by an undisclosed customer's three-year lease for 1,000 robots · subject to contract milestones · the company states that this amount is not a measure of current-period revenue · Analyst & Investor Day scheduled for 2026.10.06 · S-4 · Financial coverage
39 · UBTECH · H1 2026 results · total revenue RMB 1.269bn · full-size embodied-intelligence humanoid revenue RMB 590.3m, or 46.5% of total revenue · 921 units sold · +1,445% YoY · because this revenue includes RMB 21.247m from education use and RMB 569m from customized industrial and commercial solutions, it is not treated as revenue from autonomous labor at industrial sites · Article
40 · UBTECH · Liuzhou factory · began operation in 2026.09 · designed for one unit every 10 minutes · planned annual production capacity 10,000 units · gross floor area 14,000 m² · evidence of manufacturing capacity and factory operation · not evidence of 10,000 annual sales or 10,000 autonomous units operating at customer sites · “one unit every 10 minutes” and “10,000 annual capacity” cannot be converted into one another because the operating schedule is not disclosed · Article
41 · 2027 Budget Proposal · announced 2026.09.10 · KRW 62.0bn for sensors, actuators and secondary batteries for humanoids · 5.3× the KRW 11.6bn budget in 2026 · a budget proposal, not disbursed funding · the amount covers three component categories, not the entire robotics budget · Article
42 · Nscale · Figure · strategic partnership 2026.09.03 · up to 100,000 NVIDIA GPUs · $3.5bn initial compute commitment · intent to expand beyond $6bn · first deployment targeted for H2 2027 · Nscale becomes a Figure shareholder · a compute commitment, not cash raised · and does not mean 100,000 GPUs are currently running · Original
43 · China Humanoid Robot & Embodied Intelligence 100-Person Forum · “2026 Humanoid Robot Industry Development Report” 2026.08.20 · China H1 2026 shipments above 40,000 units · 97% of global shipments · the report text does not define “humanoid” or the point at which a unit is counted as shipped · definitions therefore differ from Counterpoint's 22,000-plus global figure released the same day · Article
44 · Rodney Brooks · “Why Today's Humanoids Won't Learn Dexterity” 2025.09.26 · argues that human dexterity depends heavily on tactile receptors and that human-action video lacks this information · “collecting only visual data is not collecting the right data” · a personal essay, not an academic paper · Original
45 · NVIDIA · Jensen Huang · GTC Taipei 2026 keynote 2026.06.02 · “the biggest problem is data” · “data for AI robots has to be from the robot's point of view, while most of the world's video data is third-person rather than first-person” · “in physical AI, compute becomes data” · describes bootstrapping with teleoperated demonstrations, simulation and synthetic data · NVIDIA is a direct participant in this market · Full transcript
46 · Rainbow Robotics · company press release, “MOU Signed with CJ Logistics to Jointly Develop AI Humanoid Logistics Robots” · a joint-development MOU stage, not a record of delivery or field validation · Original · CJ Logistics press release
This research explains the commercialization stage and economics of the humanoid industry · it is not a recommendation to buy or sell any security or asset · company claims are distinguished from customer confirmations and regulatory filings · some evidence relies on company-provided assumptions, field trials and media reports · national labor-cost figures come from different statistical series and are used only for directional comparison · operating hours, human intervention, failure intervals and investment returns not supported by public evidence are not estimated.
Research No.06 · Published 2026.09.19 · Public-source review cutoff 2026.09.19 · www.snowshagal.com
