Back to the comparison
DownloadOpenAI “deep research”

Commercial Humanoid Robots: Field Deployment, Unit Economics, and Technical Reality, 2022–2026

Executive summary

As of August 31, 2026, the public evidence does not support the often-implied narrative that thousands or tens of thousands of humanoid robots are already doing economically productive work in factories and warehouses. Roughly 20,000 humanoids were shipped globally in 2025, according to BofA data reported by Reuters, and individual Chinese manufacturers report thousands of sales, but shipments overwhelmingly include research systems, government procurement, training/data-collection fleets, demonstrations, entertainment, and industrial pilots. Reuters' examination of hundreds of filings and procurement records concluded that widespread commercial demand remains limited and that many supposed industrial deployments are still validation or training programs. [1][2]

My strict deployment ledger counts a robot only when there is evidence that it is physically operating in a normal commercial workflow and generating economic output, rather than proving a concept, collecting training data, or demonstrating technology.

Under that definition:

For bipedal humanoids, the public record establishes only a single OEM—Agility Robotics—with clearly documented, ongoing commercial customer operations at meaningful scale. Agility reports eight commercial deployments, three additional booked deployments, and current commercial deployments involving GXO, Schaeffler, Toyota Motor Manufacturing Canada, Amazon, and Mercado Libre. The company does not disclose how many Digit robots those deployments contain. A conservative inferential floor is therefore at least five simultaneously deployed Digit units, one per named commercial customer organization, but the real number is almost certainly higher. The SEC materials should not be read as saying “eight robots”: they say eight deployments, not eight units. [3][4]

If wheeled humanoids/mobile manipulators are included, Galbot materially changes the answer. A Chinese government source reported in May 2026 that Galbot humanoids were operating inside more than 120 retail stores and in smart pharmacies across dozens of cities, performing product retrieval, medicine picking, order processing, packaging, and customer hand-off around the clock. By August 31, Galbot said its mainland network had expanded to about 200 outlets in 50 cities. Because the official May wording establishes robots inside more than 120 stores, >120 physical Galbot units is a defensible lower bound; the exact current fleet and robot-per-store ratio are undisclosed. Galbot uses a wheeled base rather than humanlike legs, so it should be separated from strict biped counts. [5][6]

That yields the most defensible public lower bounds:

Definition Defensible public minimum as of Aug. 31, 2026 What is actually known Confidence
Strict bipedal humanoids doing paid production work ≥5 units Five named Agility commercial customer organizations; eight deployments reported, robot count withheld Medium: unit floor is inferred; commercial status is High
Broader humanoid form factor, including wheeled humanoids >125 units >120 Galbot robots implied by >120 robot-equipped stores, plus ≥5 Digit units Medium-High
Likely current broader installed productive fleet Order of 200+, not auditable Galbot alone says ~200 outlets by Aug. 31; exact robots/site unknown, and Digit fleet undisclosed Medium-Low
2025 global humanoid shipments ~20,000 Includes research, training, demos, government purchases, pilots and productive units Medium-High, but not a deployment figure

The gulf between ~20,000 shipments and only a few hundred publicly evidencable productive installations is the central finding of this research. [1][5][6]

Economically, the most transparent Western commercial example is Digit. Agility says its current Digit v4 bill of materials is approximately $125,000, with a long-term roadmap toward ~$30,000. Its illustrative RaaS model generates approximately $500,000 over five years per deployed robot; an ownership-plus-software/support model generates roughly $400,000. Agility's internal deployment cost is about $15,000 and its cost to deliver software and maintenance is about $15,000 per year, excluding R&D and SG&A. [7][8][9]

At today's $125,000 BOM, that implies approximately $215,000 of direct five-year lifecycle delivery cost before R&D, SG&A, financing, exceptional repairs and warranty reserves: $125,000 hardware + $15,000 deployment + $75,000 five-year software/maintenance delivery. Agility reported $74 million of R&D expense in 2025 and projects roughly $115 million in 2026, illustrating why fully loaded costs are dramatically higher than BOM at low production volume. [7]

China shows how quickly manufacturing economics can change. Unitree's IPO prospectus reports humanoid sales rising from 5 units in 2023 to 410 in 2024 and 3,551 in January–September 2025, while average realized price fell from RMB 593,400 to RMB 260,700 to RMB 167,600. Its humanoid gross margin was still 62.91% in January–September 2025, implying average manufacturing COGS of only about RMB 62,163 per humanoid for that sales mix. That is a real audited commercial cost signal—but it should not be interpreted as the cost of an autonomous industrial worker: Unitree explicitly spans research, education, entertainment, consumer and industry markets, and its prospectus describes much industrial humanoid deployment as still being in exploratory validation. [10][11][12][13][14]

The cost stack remains dominated by motion hardware. McKinsey estimates actuators at 40–60% of hardware BOM, sensing/perception at 10–20%, compute/control at 10–15%, structure at 5–10%, and batteries at 5–10%. [15]

The most rapidly improving problems since 2022 have been locomotion, hardware price, actuator performance, design-for-manufacture, narrow-task manipulation reliability, and charging/duty-cycle engineering. The problems improving much more slowly are the ones that determine economic usefulness: generalization to exceptions, dexterous manipulation of varied objects, predictable uptime and MTBF, speed relative to purpose-built automation, safe fenceless operation around humans, and the data burden needed to make robots robust outside trained workflows. Reuters found that conventional industrial arms still generally outperform humanoids in uptime and reliability, while industry participants describe current humanoids as too slow and error-prone for broad adoption. [1][2]

The practical conclusion is that the sector has crossed from “pure prototype” into real but narrow commercial deployment, not into mass labor substitution. The strongest economics today occur where the task is repetitive, structured, ergonomically unpleasant, does not require high dexterity, and where a humanoid can reuse human-oriented infrastructure without a costly line redesign.

Scope, definitions, and evidence standard

This report covers developments from 2022 through August 31, 2026, with emphasis on the United States, Canada, China and Europe. It distinguishes four stages that are too often conflated in company announcements:

Classification Test used in this report Counted as genuine paid field deployment?
Research/demo Lab testing, trade-show work, dancing, teleoperated training, publicity demonstrations No
Proof of concept Robot installed temporarily to see whether a workflow is feasible No
Pilot Customer-site operation explicitly described as a pilot, validation program or data-collection exercise, even when it touches real products No
Commercial production deployment Ongoing customer operation after validation, under RaaS, purchase/service agreement, or an internal revenue-generating business workflow Yes

That distinction matters because company terminology is inconsistent. Apptronik, for example, called its Mercedes relationship a “commercial deployment” in a headline while the same official announcement explicitly says the agreement is to pilot Apollo and identify potential applications. It is therefore excluded. [16] Figure's BMW program likewise handled actual production components and contributed to customer vehicles, yet BMW itself calls it a pilot project, so it is excluded from the paid-deployment count. [17][18]

Likewise, a sale is not automatically a productive deployment. Unitree's prospectus shows very large, legitimate humanoid sales volumes, but also says its platforms span scientific research, education, commercial consumption and industry, and notes that many industrial humanoid programs remain in exploration and validation. [14] UBTECH shipped 1,079 full-size humanoids in 2025, but Reuters found that state-linked contracts, training centers and testing comprise a significant part of Chinese demand; a major UBTECH government installation visited by Reuters was explicitly collecting teleoperated training data rather than replacing production labor. [1][2]

Confidence levels in the tables mean:

High: primary filing, regulator, customer, procurement record or official source gives the number or status directly.
Medium: primary data establishes most of the fact, but a small inference is required—for example, one robot must exist in each concurrently robot-operated store.
Low: management target, analyst estimate, secondary source without enough underlying disclosure, or a derived figure sensitive to assumptions.

The cut-off date matters in one instance: Galbot held the opening ceremony for three Hong Kong robot-operated stores on August 31, but SCMP says the stores were to launch the following day, September 1. Those Hong Kong locations are therefore not counted as operating field deployments at this report's cut-off. [6]

Genuine commercial deployments

The strict deployment ledger

Company / model Form factor Verified qualifying deployment Verified units or defensible floor Primary customers / operators Main capabilities in production Price / commercial model Confidence and evidence
Agility Robotics — Digit v4 Biped Yes Exact fleet ND; ≥5 inferential floor; company reports 8 commercial deployments + 3 booked, with 9 customer-facility commitments GXO, Schaeffler, Toyota Motor Manufacturing Canada, Amazon, Mercado Libre Tote movement, tote fulfillment/handling, material movement and logistics RaaS or ownership; illustrative ~$500k/5 yr RaaS or ~$400k/5 yr ownership+software/support High for commercial status/customer list; Medium for ≥5-unit floor because units are not disclosed. SEC materials explicitly call these commercial deployments. [3][9]
Galbot — G1 Wheeled humanoid / dual-arm mobile manipulator Yes >120 physical-unit lower bound by May 2026; network ~200 outlets by Aug. 31, exact fleet ND Galaxy Space Capsule retail network; smart-pharmacy operators; end consumers Order taking, product recommendation, shelf retrieval, medicine identification/picking, packaging, hand-off; food/drink service in some sites Enterprise quote; engineer cited ~RMB 700,000 in early 2026; commercial stores also monetize retail transactions High for >120 robot-equipped stores and tasks; Medium for translating sites into a minimum unit count; current ~200 is a site count, not unit disclosure. [5][6][19]
Figure — Figure 02 Biped No — excluded pilot 0 in strict ledger BMW Spartanburg Sheet-metal part loading into weld fixtures; high-precision pick/place Price ND High. Figure reports 1,250+ operating hours and 90,000+ parts, but BMW explicitly calls the program a pilot. [17][18]
Unitree — H1/G1/R-series Biped Large commercial sales, but productive field fleet not separable 5,500+ humanoids shipped in 2025, but not counted as 5,500 workers Universities/research institutes, developers, consumers, commercial and industry users Locomotion, research, education, performance and emerging industrial manipulation G1 base price RMB 85,000 in 2026 prospectus; Jan–Sep 2025 realized humanoid ASP RMB 167,600 High for sales; High that shipment count cannot be equated to productive deployment. [20][10][14]
UBTECH — Walker S1/S2 Biped Commercial hardware sales are real; productive-unit count not auditable 1,079 full-size humanoids shipped in 2025, but strict productive fleet ND State-owned enterprises; automotive partners including testing programs Material handling, industrial manipulation; S2 autonomous power management / battery swapping List price ND; 2025 humanoid revenue $122m versus 1,079 shipments gives only a crude portfolio revenue-per-unit proxy High for shipments/revenue; Medium-High for exclusion from strict fleet because Reuters found significant state/training demand and factory testing rather than production substitution. [2][21]
Apptronik — Apollo Biped No — pilot 0 strict Mercedes-Benz; Jabil and other validation partners Kitting, delivery of totes/components, inspection concepts ND High. Apptronik's own announcement says “pilot” and “exploring” applications. [16]
Tesla — Optimus Biped No verified external paid deployment 0 strict external units Tesla internal development Training/data collection and development toward manufacturing tasks Not commercially sold Medium-High. Unitree's regulated prospectus notes Optimus is not externally sold and was primarily internal development/testing; Tesla's public page still frames it as a robot it is developing. [14][22]
Boston Dynamics — Atlas Biped No verified productive customer operation at cut-off 0 strict 2026 fleets committed to Hyundai RMAC and Google DeepMind, with broader customers planned later Material handling/order fulfillment target; autonomous battery swap; 50 kg lift ND High for status: Boston Dynamics announced manufacturing and scheduled deployments but simultaneously describes Atlas as “currently in development”; no production KPIs or paid-work utilization have been published. [23]

“ND” means not disclosed.

The result should not be read as “only five bipedal humanoids exist in factories.” It means that five is the defensible named-customer unit floor that can be supported without turning site counts, shipments or pilots into imaginary production fleets. Agility itself clearly has more: it says eight commercial deployments and reports 65,000 hours of operation associated with deployment commitments, but its public filings do not reveal robot counts at each site. [4][3]

Agility's evidence is unusually strong because GXO and Agility announced in June 2024 that Digit had moved from testing into a multi-year RaaS arrangement and was integral to day-to-day operations at GXO's Georgia facility. By November 2025, Agility reported that Digit had moved more than 100,000 totes there. [24][25][26] Toyota Motor Manufacturing Canada subsequently moved from a successful pilot into a commercial RaaS agreement in February 2026. [27]

Galbot is the other important exception to the pilot-heavy pattern, although its wheeled morphology merits a separate category. The China National Intellectual Property Administration reported robots operating inside more than 120 retail stores, with smart-pharmacy deployments in dozens of cities, explicitly describing them as “24/7 sales assistants” handling normal store orders. By August 31, Galbot reported about 200 outlets across 50 mainland cities. [5][6]

Publicly evidenced commercial deployment sites; site/deployment count is not the same as robot-unit count

The chart deliberately shows sites/deployments rather than claiming robot fleets that companies have not disclosed. Galbot's >120 May figure is stronger than a generic site count because the government source specifically states that humanoid robots are inside those stores; Agility's “eight deployments” is left as a deployment measure because its SEC deck does not give robots per deployment. [5][3]

Why the shipment numbers are misleading

Unitree demonstrates the scale-up vividly. Its regulated prospectus reports only 5 humanoids sold in 2023, 410 in 2024, and 3,551 in the first nine months of 2025; the Shanghai Stock Exchange later reported that full-year 2025 humanoid shipments exceeded 5,500. [11][20]

Unitree humanoid sales reported in its IPO prospectus

Yet Unitree's own competitive discussion says industry applications of humanoids remain substantially in exploration and validation, while its products are widely used by universities, research institutions, developers and commercial/consumer users. [14] That distinction explains how thousands of genuine robot sales can coexist with a much smaller population of robots autonomously earning their keep on production tasks.

Reuters found the same pattern at industry level. About 20,000 humanoids shipped worldwide in 2025, approximately 95% from Chinese manufacturers according to BofA, but Chinese public-sector entities also spent at least $230 million in the first half of 2026 on robots, training systems, equipment and demonstration projects. Reuters inspected nearly 1,000 fulfilled tenders and concluded that state purchases and subsidies were driving much of current demand. [1]

UBTECH is a useful caution. It shipped 1,079 full-size humanoids in 2025 and won at least $123 million of new Q4 contracts with state-owned enterprises—more than its entire 2025 humanoid revenue of $122 million. But Reuters' factory reporting found several automotive programs were tests, while a >100-robot government-backed facility it visited was using teleoperators to generate embodied-AI training data. Those robots are commercially sold hardware, but counting them as autonomous workers would answer a different question from the one asked here. [1][2]

Daily work, utilization, and operating performance

The work that has crossed the commercial threshold is strikingly narrow. It is mostly pick, carry, retrieve, place, and hand off in relatively structured environments.

Model / deployment What the robot actually does Actual operating evidence Utilization / duty-cycle evidence Interpretation Confidence
Digit at GXO Picks/moves totes between autonomous mobile-robot interfaces, conveyors and human-oriented logistics areas >100,000 totes at GXO by Nov. 2025; Agility cites ~98% task accuracy in deployment proof points Exact uptime not published. Agility's economic model assumes 120 operating hours/week, equivalent to 71% of all calendar hours Strongest biped commercial benchmark; narrow material-handling workflow High for task/output; Medium-Low for utilization because 120 h/week is a model, not measured uptime. [26][3][28]
Digit broader fleet Tote fulfillment/handling, material movement, logistics Company reports 65,000 hours of operation, eight commercial deployments and 9 facility commitments Digit v4 hardware: ~4 h runtime plus autonomous charging; v5 targets substantially higher duty cycle, up to ~22 h/day of battery-supported output Real operational learning, but fleet count and commercial/pilot split of 65k hours are not disclosed High for company figures; Medium for commercial utilization interpretation. [4][29]
Galbot G1 retail/pharmacy Processes online orders; recognizes/retrieves merchandise and medication; packages and hands items over; some sites serve food/drinks >120 stores by May; dozens of cities for pharmacy; Reuters reports roughly one minute to retrieve/prepare an order and >95% success in observed pharmacy-class workflows Service proposition is 24/7; actual active-motion percentage, MTBF and intervention minutes/day are undisclosed Economically useful mainly because stores need coverage over long hours; robot need not beat a human on instantaneous speed High for operating window/tasks; Medium for performance metrics. [5][30]
Figure 02 at BMW — excluded pilot Loads sheet-metal components into fixtures for vehicle body production 90,000+ parts, 1,250+ runtime hours, contributed to 30,000+ X3 vehicles over an 11-month deployment Scheduled around a 10-hour shift Monday–Friday; exact number of robots and fleet uptime not provided Excellent evidence that narrow manipulation is becoming industrially credible, but BMW calls it a pilot High. [17][18]
Walker S2 — technical benchmark, not counted fleet Intended for factory handling/manipulation No independently auditable paid-work fleet utilization number Dual batteries and station integration allow robot to choose charging or swapping; 15 kg payload Solves power continuity mechanically, but “can work continuously” is not the same as measured production uptime High for hardware functionality; Low for real utilization. [21]

Agility's modeled 120 hours per week deserves interpretation. That is the equivalent of six 20-hour days or another scheduling pattern totaling 120 hours, and equals about 71.4% utilization of the 168-hour calendar week. It is a customer-economics assumption, not an audited measured average. Agility's deck compares that robot service with fully burdened human labor priced at $30.50/hour and assumes two ten-hour shifts. [4]

The economics can therefore work even without human-level speed if the robot can cover shifts that are expensive or hard to staff, sustain adequate reliability, and avoid major facility redesign. That appears to be the logic behind both Digit's tote work and Galbot's overnight retail/pharmacy use. Agility explicitly targets structured, repetitive workflows in environments already designed around people. [3]

Conversely, throughput remains a major weakness. Reuters' investigation found that conventional robotic arms generally outperform humanoids on uptime and reliability, and industry participants said existing automotive automation is faster and easier to control. Training-center humanoids were substantially slower than human workers on simple tasks, while failures often emerge when object pose, lighting, geometry or other details deviate from trained conditions. [1][2]

A key implication is that “utilization” has at least three different meanings that vendors sometimes blur:

Scheduled availability is the time a robot is assigned to a shift.
Powered availability is time it can stay operational given charging/swapping.
productive utilization is time performing successful value-adding cycles rather than charging, waiting, recovering, being teleoperated or undergoing maintenance.

Public humanoid disclosures increasingly provide the first two, but almost never provide the third as an audited percentage. That is one of the biggest data gaps in the sector.

Unit economics: manufacturing, R&D, integration, support, and pricing

What “cost per humanoid” actually means

A quoted “$20,000 robot” or “$125,000 BOM” is not the same thing as the economic cost of delivering five years of reliable labor. A useful decomposition is:

Fully loaded robot economics = manufacturing/BOM + assembly/test + deployment/integration + software/cloud + support/maintenance + spares/warranty + R&D amortization + corporate overhead + financing/capital carrying cost.

For a RaaS vendor, the company also carries asset-life and utilization risk. If a robot spends months idle, the customer may be insulated while the vendor still owns a depreciating machine.

Digit: the best disclosed end-to-end example

Agility provides enough disclosure to construct a reasonably grounded economic stack.

Digit cost/revenue item Public figure Nature of figure Confidence
Current Digit v4 BOM ~$125,000 Current manufacturing BOM stated by management, June 2026 High [8]
Long-run BOM objective ~$30,000 Cost roadmap, not achieved cost Medium-Low / forward-looking [8]
Agility's one-time deployment cost ~$15,000 Company's internal delivery cost High [7]
Customer deployment fee, RaaS model ~$25,000 Illustrative customer charge High as model; not necessarily every contract [7]
Customer deployment fee, ownership model ~$20,000 Illustrative customer charge High as model [7]
Cost to deliver software + maintenance ~$15,000/year Vendor cost High [7]
Five-year RaaS revenue ~$500,000 Illustrative cumulative revenue per robot High as management model, not guaranteed realized revenue [9]
Five-year ownership + software/services revenue ~$400,000 Illustrative cumulative revenue High as model [9]
2025 R&D expense ~$74m Company actual High [7]
2026 R&D planning assumption ~$115m Forecast/model assumption Medium [7]
2026 SG&A planning assumption ~$60m Forecast/model assumption Medium [7]

At today's v4 BOM, a simple five-year direct-cost calculation is:

$125k hardware + $15k deployment + 5 × $15k software/maintenance = ~$215k direct five-year cost.

That does not include R&D, SG&A, financing, unusual repairs or warranty losses, consistent with Agility's own disclosure that those corporate costs sit outside its illustrated product-cost calculation. [7]

R&D amortization illustrates why production scale is decisive. Using Agility's approximately $115 million 2026 R&D assumption purely as a sensitivity analysis:

Annual production scale over which $115m R&D were spread Illustrative R&D burden per unit
500 $230,000
1,000 $115,000
5,000 $23,000
10,000 $11,500

Those are not Agility forecasts or accounting allocations; they show why a $125,000 BOM can coexist with a business that still needs hundreds of thousands of dollars of lifetime revenue per unit. The same logic applies to SG&A and integration engineering. Agility itself reports a 10,000-robot annual capacity at RoboFab, but capacity should not be confused with current output. [4]

At the 1,000-unit illustrative scale, adding $115,000 of R&D allocation and $60,000 of SG&A allocation to the $215,000 direct five-year burden produces roughly $390,000 per unit before financing and exceptional service costs. At 10,000 units, those corporate allocations shrink by a factor of ten. Meanwhile, the BOM itself is also expected to decline with volume, so this should be read as scale sensitivity rather than a static cost forecast.

Agility's RaaS pricing is designed around customer labor economics rather than a simple hardware markup. Management says the vast majority of more than $300 million of contracted Digit v5 orders are structured as RaaS, and the company prices the service at a discount to fully burdened human labor. [4][8][9]

Unitree: what high-volume Chinese manufacturing reveals

Unitree's audited prospectus gives a useful counterpoint. For January–September 2025 it reports:

That mathematically implies average COGS of:

RMB 167,600 × (1 − 0.6291) ≈ RMB 62,163 per unit.

This is a derived estimate from audited sales and margin data, not a disclosed component BOM. It likely incorporates the manufacturing-cost accounting used in Unitree's gross margin and reflects a product mix skewed toward lower-cost G1-class units, not a fleet of fully integrated autonomous factory workers.

Unitree's company-wide January–September 2025 production cost structure was 80.12% direct materials, 8.04% direct labor and 11.85% manufacturing overhead, emphasizing how strongly robot economics depend on component sourcing and design-to-cost. [31]

R&D was RMB 90.21 million for the same nine-month period, or 7.73% of revenue. If one makes the explicitly simplifying assumption that humanoid products consume R&D in proportion to revenue, then 7.73% of the RMB 167,600 average selling price represents approximately RMB 12,950 of R&D burden per humanoid sold. That proportional allocation is analytical rather than company reporting. Unitree's prospectus shows actual R&D concentrated in employee compensation, prototype/direct materials and, increasingly, cloud compute for large-model development. [32][31]

The contrast with Agility should not be interpreted as “Unitree can build the same industrial worker for one-tenth the price.” Their product mix and customer use are radically different. Unitree's volume includes research, education and consumer/commercial use; Digit's economics include a managed industrial workflow and ongoing service. [14][9]

Industry-level cost trajectory

Goldman Sachs reported in early 2024 that humanoid material costs had fallen about 40% faster than its previous expectations, causing it to materially raise shipment forecasts. [33] Morgan Stanley estimated that a humanoid in a high-income-country cost environment cost roughly $200,000 in 2024, projecting approximately $150,000 by 2028 as technology and scale improve. [34]

Chinese price compression has been much faster. Unitree's realized average humanoid selling price fell from RMB 593,400 in 2023 to RMB 260,700 in 2024 and RMB 167,600 in January–September 2025—a decline of roughly 72% from 2023 to the 2025 nine-month period. [10] Reuters reported that Leju's full-size Kuavo average price fell 26% year over year in 2025 to about $46,000, and Morgan Stanley expected average humanoid prices to fall another 15% during 2026. [2]

These declines are evidence of real manufacturing progress. They are not yet evidence of proportional reductions in cost per productive labor-hour, because autonomy, support, intervention and utilization remain bottlenecks.

Where the costs concentrate

McKinsey's April 2026 supply-chain analysis gives the clearest current top-level hardware breakdown. It estimates the following ranges: [15]

Hardware category Share of humanoid BOM Why it is costly
Actuators / joint drive systems 40–60% Many high-torque, compact, backdrivable or high-efficiency joints; motors, reducers, encoders, drives and thermal requirements multiply across dozens of axes
Sensing and perception 10–20% Cameras, depth systems, force/torque sensing, joint sensing, tactile sensing and perception electronics
Compute and control 10–15% Edge AI compute, real-time motion-control electronics, power/control boards and communication
Structure 5–10% Lightweight stiff frames, housings, precision parts and impact-resistant assemblies
Battery modules 5–10% High-discharge cells, battery management, packaging, thermal and rapid-swap architecture
Other Remainder Wiring harnesses, cooling, connectors, miscellaneous electronics and hardware

McKinsey notes that hands can either be treated as their own subsystem or embedded in the actuation and sensing categories; economically, they remain particularly important because useful hands combine many small actuators, transmission components and sensors in a tight volume. [15]

Illustrative point allocation within McKinsey's published BOM ranges; this is not an audited industry-average pie

The pie is deliberately an illustrative feasible allocation—45% actuators, 12% sensing, 12% compute/control, 9% structure, 10% batteries and 12% other—chosen to sit within McKinsey's published ranges. The published ranges themselves are the evidence; no industry-wide audited “exact percentage” exists. [15]

The importance of actuators is corroborated by OEM strategy. Unitree says it vertically develops key motors, reducers, integrated joints, dexterous hands, lidar and sensing, and specifically identifies motor–reducer–drive–sensor–thermal integration as a core cost and performance advantage. [35] Boston Dynamics has likewise arranged for Hyundai Mobis to supply Atlas actuators and describes actuator development and scalable supply as central to its production strategy. [23]

There are therefore two complementary ways to view the cost stack:

By component: joints/actuators dominate, followed by sensing and compute. [15]

By manufacturing input: at a scaled Chinese OEM such as Unitree, materials comprise about 80% of reported production COGS, while direct factory labor is only about 8%. [31]

That combination explains why supply-chain localization, actuator integration, lower part count and design for high-volume processes have generated such large cost declines. Factory labor is not the major cost to squeeze; components and engineering complexity are.

Software is different. Software usually has relatively low marginal replication cost, but the development cost is enormous. Unitree's 2025 nine-month R&D expense was dominated by employee compensation and included rapidly growing cloud-compute expenditure; Agility spent $74 million on R&D in 2025 despite a tiny installed fleet relative to conventional industrial automation. [31][7] Thus software's percentage of BOM may be nearly zero while its percentage of fully loaded economic cost at low volume is very large.

Technical progress versus persistent bottlenecks

The evidence from 2022–2026 suggests a bifurcation: body hardware has improved faster than general-purpose physical intelligence.

timeline
    title Humanoid robotics: what moved from 2022 to 2026
    2022 : Most current platforms remain prototype-stage
    2023 : Unitree sells 5 H1 humanoids at high early-production pricing
    2024 : H1 reaches 3.3 m/s class locomotion
         : Humanoid material costs fall sharply
         : Digit enters first GXO commercial operation
         : Figure begins BMW production pilot
    2025 : Unitree exceeds 5,500 annual humanoid shipments
         : Digit passes 100,000 totes at GXO
         : Figure 02 logs 1,250+ BMW runtime hours
         : Autonomous battery swapping appears on Walker S2
    2026 : Digit reports 8 commercial deployments
         : Galbot exceeds 120 robot-equipped stores, later ~200 outlets
         : Hardware prices continue falling
         : ISO industrial humanoid safety standard still unfinished
         : Dexterity, generalization and reliability remain key blockers

The milestones are supported by Unitree's prospectus and Shanghai Stock Exchange disclosures, Agility's filings and commercial case studies, Figure/BMW production reporting, UBTECH's technical materials, Galbot deployment evidence, and ISO's live standards record. [11][20][26][17][21][5][36]

Areas that have improved rapidly

Locomotion and dynamic balance have largely ceased to be the primary obstacle for structured industrial floors. Unitree reports that H1 reached approximately 3.3 m/s in 2024 and progressed to high-dynamic maneuvers, while its 2025–2026 systems demonstrated increasingly sophisticated coordinated locomotion. These achievements do not prove useful manipulation, but they show that basic biped balance and mobility have advanced rapidly. [37]

Hardware manufacturing cost and commercial price have fallen extraordinarily quickly. Unitree's realized humanoid ASP fell from RMB 593,400 in 2023 to RMB 167,600 in the first nine months of 2025 while maintaining a reported 62.91% humanoid gross margin. Goldman Sachs separately reported a roughly 40% reduction in material cost around 2024 versus its earlier expectations. [10][13][33]

Design for manufacturing has moved from prototype machining toward scalable production engineering. Unitree reports mass-production-stage integrated-joint technology and flexible production of motors, reducers, batteries, electronics and structural components; Agility operates a facility designed for 10,000 Digit units/year; Boston Dynamics' 2026 Atlas redesign explicitly reduced unique-part count and was designed around automotive supply chains. Production capacity is not production volume, but the manufacturing architecture is visibly maturing. [35][4][23]

Narrow, repetitive manipulation is becoming reliable enough for real workflows. Digit has exceeded 100,000 tote movements in a live GXO operation and Agility reports approximately 98% task accuracy in deployment proof points. Figure's excluded-but-informative BMW pilot executed more than 90,000 production-part placements over 1,250 hours and operated around automotive tolerances. Galbot reports commercial retail/pharmacy picking at high success rates. [26][3][17][30]

Power continuity has improved, mostly by systems engineering rather than a battery-energy-density revolution. Digit v4 combines roughly four-hour runtime with autonomous charging; the forthcoming v5 is designed for materially higher daily battery-supported output. Walker S2 uses dual batteries and autonomous station interaction so it can choose between swapping and charging according to workload. Boston Dynamics' production Atlas is similarly designed to change its own batteries. [4][29][21][23]

That distinction matters: battery life itself remains limited, but vendors have improved duty cycle through automated charging, hot swapping and scheduling.

Areas that remain stagnant or stubbornly difficult

Generalization is the central software bottleneck. Reuters' factory and training-center reporting found humanoids struggling when tasks deviate from learned routines. Industry experts characterized the problem as becoming dramatically harder as required placement precision narrows from centimeters toward millimeters. Hardware can walk to the station; deciding correctly what to do when an object is misplaced, partly occluded, deformed, upside down or unexpectedly blocked remains far less robust. [1]

Dexterity and contact-rich manipulation remain among the hardest subsystems. McKinsey explicitly identifies hands as a persistent technical challenge despite rapid progress elsewhere, because economically relevant picking, sorting and tool use require reliable manipulation of diverse objects. Reuters independently reached the same conclusion after factory visits. [15][1]

Figure's BMW pilot illustrates both progress and remaining fragility. It achieved meaningful industrial runtime, but Figure reported that the forearm was the leading hardware failure point, prompting a wrist/electronics redesign in its next generation. [17]

Human-equivalent speed remains elusive. Purpose-built industrial automation typically wins on cycle time, repeatability and uptime in stable high-volume manufacturing. Reuters found that automotive operators still view conventional robotic arms as faster and easier to control and that humanoids remain too slow and error-prone for broad production deployment. [1][2] This is why early humanoid economics focus on jobs where traditional automation is awkward—not on displacing an already optimized six-axis robot.

Measured uptime and MTBF remain opaque. Companies increasingly publish cumulative hours, tote counts and impressive shift demonstrations, but almost none publish standardized field metrics such as mean time between failures, percentage of scheduled hours completed autonomously, human intervention minutes per operating hour, or cost per successful cycle. Agility's 65,000 operating hours and Figure's 1,250-hour BMW pilot are valuable, but without fleet size and downtime denominators they do not produce a comparable utilization percentage. [4][17]

Safety for fenceless human collaboration is not mature as a standards regime. Agility's own SEC materials acknowledge that current humanoid deployment generally requires separation of humans and robots and position v5 as a step toward cooperative safety. [38] More importantly, the relevant ISO work remains unfinished: as of August 31, ISO 25785-1 is only a Committee Draft under development, covering dynamically stable industrial mobile robots including bipeds. [36]

There is a noteworthy source-integrity discrepancy here. An Agility investor deck characterizes ISO 25785-1 as a new approved safety standard, whereas ISO's own live catalog lists it as a Committee Draft and shows the standard still progressing through development stages. For standards status, the ISO record is authoritative, so this report treats the standard as not yet published. [29][36]

Training-data scarcity remains structural. Chinese governments and companies are building large teleoperation centers precisely because there is no internet-scale corpus of high-quality robot interaction trajectories comparable with text or images. Reuters observed >100 humanoids in one subsidized center being teleoperated to create training data, and the facility itself acknowledged that its current economic model had no clear path to profitability. [1]

Integration remains expensive and task-specific. Even Agility, which has one of the industry's most mature commercial workflows, budgets a one-time deployment cost and moves customers through technology proof, on-site proof of concept and RaaS pilot phases before scaled deployment. Its commercial process implicitly demonstrates that a “general-purpose” robot still requires significant workflow engineering. [3][7]

The net technical picture is therefore:

Technical problem 2022–2026 progress Current status Evidence confidence
Basic biped locomotion / balance Very rapid Good enough for structured floors; spectacular dynamic motion demonstrated High [37]
Hardware BOM / manufacturing cost Very rapid Falling by tens of percent per generation/year at scaled Chinese OEMs High [10][33]
Production engineering / part count Rapid Automotive-style DFM entering next-generation products High [23][35]
Autonomous charging / battery swap Rapid systems progress Long duty windows mechanically achievable, even though battery energy remains constrained High [21][23]
Narrow pick/place reliability Rapid Commercially viable in a few structured workflows High [26][17]
General-purpose reasoning / exception handling Improving, but far behind expectations Major blocker to autonomous labor High [1]
Dexterous hands / varied objects / tool use Slow relative to locomotion Major unsolved subsystem High [15][1]
Human-equivalent task speed Limited Frequently below humans and below fixed automation High [1]
Field reliability / standardized MTBF Insufficient evidence Probably improving, but vendors disclose too little to quantify Medium
Fenceless safety / certification Slow institutional progress ISO standard still under development High [36]
Training data for broad physical intelligence Massive investment, unresolved Teleoperation and synthetic/data-center buildout still needed High [1]

Outlook, data gaps, and conclusions

The commercial humanoid sector in 2026 is best described as an emerging automation category with a very small productive installed base sitting underneath a much larger hardware-sales and training ecosystem.

China is clearly ahead in manufacturing volume and price compression. Unitree alone exceeded 5,500 humanoid shipments in 2025, while UBTECH shipped 1,079 full-size units; Reuters says Chinese producers represented roughly 95% of about 20,000 global humanoid shipments that year. [20][2][1] Yet that volume has not translated proportionally into autonomous labor. Reuters found demand heavily influenced by public procurement, subsidies and data-collection programs, and Unitree's own prospectus says industrial deployment remains substantially exploratory. [1][14]

North America currently provides the clearest evidence of bipedal robots sold as labor services rather than machines for experimentation, primarily through Agility. Digit has crossed the crucial line from “pilot” to ongoing operational RaaS, but Agility's fleet remains small enough that it withholds—or has not needed to report—robot-level installation counts. [25][3] Figure's BMW work shows that competing systems are technically close, but BMW's explicit pilot terminology keeps those robots outside the strict commercial ledger. [18] Apptronik remains in the same pilot category. [16]

Europe has customer participation through companies such as Schaeffler and Mercedes-Benz, but the same pilot-versus-commercial distinction applies: Schaeffler is included in Agility's current commercial-deployment list, while Mercedes' Apollo work is explicitly exploratory/pilot. [3][16]

The first large-scale productive use cases may not look like science-fiction “general-purpose workers.” Galbot's strongest evidence comes from a wheeled robot in compact retail/pharmacy environments. Digit's comes from carrying totes. That is economically logical: wheels are more efficient than legs on flat surfaces, and a robot that performs one repetitive logistics function reliably can create value long before it can clean a house, repair equipment or improvise with arbitrary tools.

Three economic thresholds will determine when today's few-hundred-scale evidence becomes a genuine installed base of tens of thousands.

First, cost per productive hour must fall, not merely BOM. A $30,000 robot that requires frequent human intervention can be less economical than a $125,000 machine delivering dependable autonomous shifts.

Second, autonomous utilization must become measurable and comparable. The industry needs metrics analogous to industrial automation: uptime, MTBF, mean time to recover, autonomous cycles between interventions, intervention labor per robot-hour and lifetime maintenance cost.

Third, workflow breadth must grow without proportional integration expense. A humanoid only earns the “general purpose” premium if a second or third task can be added mainly in software instead of requiring another months-long integration project.

The largest unresolved data gaps are therefore not robot top speed or degrees of freedom. They are commercially mundane: exact active fleet counts, autonomous hours per week, intervention rates, maintenance labor, spare-parts consumption, warranty expense, contract churn, realized price per robot-hour, and customer ROI after integration. No major OEM currently discloses these consistently enough to construct industry-wide comparable economics.

The most important numerical conclusion is consequently a range rather than a false-precision global count:

As of August 31, 2026, public primary evidence supports a strict minimum of roughly five bipedal humanoid units in clearly named, ongoing revenue-generating customer deployments, all from Agility Robotics, while Agility reports eight commercial deployment instances but withholds robot counts. Including wheeled humanoids, the defensible productive-field minimum rises above 125 units because Galbot had robots operating in more than 120 commercial stores by May; its network had expanded to about 200 mainland outlets by August 31, suggesting—but not proving—a current productive fleet around the low hundreds. [3][5][6]

That figure is orders of magnitude below headline shipment statistics, and the difference is not a bookkeeping curiosity: it is the most important fact for understanding the current economics of humanoid robotics.

Selected source record

Source Type Principal use in this report
Agility Robotics SEC investor presentation, June 2026 Regulatory filing / primary Commercial deployments, customers, operational hours, pricing models, revenue economics, deployment pipeline. [3][4]
Agility Robotics SEC transaction/call materials Regulatory filing / primary RaaS structure, >$300m v5 contracts, ~$500k/$400k lifetime revenue, BOM roadmap. [9][8]
GXO / Agility commercial-deployment announcements Customer/OEM primary Confirmation that Digit entered day-to-day commercial operations rather than only testing. [24][25]
Toyota Motor Manufacturing Canada / Agility commercial announcement Customer/OEM primary Conversion from pilot to commercial RaaS. [27]
Unitree Robotics IPO prospectus, Shanghai Stock Exchange Audited/regulatory primary Production, sales, ASP, gross margin, manufacturing-cost composition, R&D and application mix. [11][12][13][31]
Shanghai Stock Exchange / Unitree IPO reporting Exchange / official Full-year 2025 >5,500 humanoid shipments and company financial scale. [20]
China National Intellectual Property Administration Government primary/official >120 Galbot robot-equipped stores and commercial pharmacy task descriptions. [5]
South China Morning Post, Aug. 31, 2026 Reputable press / company executive sourcing Galbot's ~200 mainland outlets across 50 cities and Hong Kong launch timing. [6]
Reuters investigation, Aug. 27, 2026 Investigative reporting based on filings, procurement records and site visits Global shipment context, Chinese state procurement, UBTECH shipments/contracts, factory limitations, dexterity, reliability, economic reality. [1][2]
BMW + Figure 02 production reports Customer/OEM primary 90k+ parts, 1,250+ hours, shift pattern, and crucially BMW's classification of the work as a pilot. [17][18]
Apptronik official Mercedes announcement OEM primary Explicit pilot status despite “commercial deployment” marketing language. [16]
Boston Dynamics Atlas product announcement, Jan. 2026 OEM primary 2026 committed deployments, autonomous battery swap, 50 kg lift, manufacturing strategy and development status. [23]
UBTECH Walker S2 product documentation OEM primary Dual-battery dynamic power management, autonomous charge/swap architecture and 15 kg payload. [21]
McKinsey, April 2026 humanoid supply-chain analysis Reputable industry analysis BOM category percentages and identification of actuators/hands as cost and supply-chain bottlenecks. [15]
Goldman Sachs Research, Feb. 2024 Reputable financial research Evidence of rapid material-cost decline and upward revision of humanoid shipment economics. [33]
Morgan Stanley Research, 2025 Reputable financial research Approximate $200k 2024 high-income-country humanoid cost and long-run price trajectory. [34]
ISO 25785-1 live catalogue Standards body / authoritative primary Confirmation that industrial dynamically stable mobile-robot safety standard is still under development rather than a published International Standard. [36]

Sources

  1. No link given.
  2. No link given.
  3. No link given.
  4. No link given.
  5. No link given.
  6. No link given.
  7. No link given.
  8. No link given.
  9. No link given.
  10. No link given.
  11. No link given.
  12. No link given.
  13. No link given.
  14. No link given.
  15. No link given.
  16. No link given.
  17. No link given.
  18. No link given.
  19. No link given.
  20. No link given.
  21. No link given.
  22. No link given.
  23. No link given.
  24. No link given.
  25. No link given.
  26. No link given.
  27. No link given.
  28. No link given.
  29. No link given.
  30. No link given.
  31. No link given.
  32. No link given.
  33. No link given.
  34. No link given.
  35. No link given.
  36. No link given.
  37. No link given.
  38. No link given.