top of page
Search

Humanoid Robots in Manufacturing 2026–2027: What BMW, Mercedes-Benz and Foxconn Prove—and What Buyers Still Need to Know

Writer: Ahtesham Shaikh
Ahtesham Shaikh
Oct 23
35 min read

Physical AI Journal — Independent analysis by Sekason Research Limited, United Kingdom | Research cut-off: 23 October 2026 · Reader: Automation/Procurement Buyer · Evidence labels: (company-claimed) marks a figure that originates with a manufacturer, OEM or vendor; (verified) marks an independently sourced figure.


Physical AI Journal cover with orange circuit graphics and headline: Humanoid Robots in Manufacturing 2026–2027.

Scope & Disclaimer

This report provides market and adoption intelligence only. It is not legal, financial, investment, engineering, safety-certification or technical-conformity advice, and it does not endorse or certify any platform or manufacturer. Figures that originate with companies are not independently verified, and forecasts are uncertain. Readers should verify claims with manufacturers and consult qualified advisers before any decision. Safety and conformity questions belong with qualified engineers.


Nothing in this report is a view on whether any robot or deployment meets a safety or technical standard; standards status is reported as published. Cost figures marked illustrative are not market data. The evidence base closes on 23 October 2026, and later developments are not reflected.

See the: Disclaimer and Editorial Standards pages for the full terms.


1. Executive Intelligence Synthesis

Run a bounded paid pilot on one narrow, repeatable task and defer production-scale capital for humanoid robots in manufacturing. BMW and Figure AI report more than 30,000 X3 vehicles supported and about 1,250 operating hours (company-claimed), but no site, including Mercedes-Benz and Foxconn, has independently verified economics. Disclosed terms and measured results, which a purchase case needs, do not exist in the public record at 23 October 2026.


This report helps Automation/Procurement Buyers decide whether to commit capital now, run a bounded paid pilot, or defer procurement until operational and commercial evidence supports wider deployment. The public record shows documented factory activity, but the figures that exist describe activity, not economic return. A buyer who treats a vendor-reported total as a measured result risks approving capital on the strength of a demonstration.


The report is organised around that distinction. Sections 2 and 3 grade the evidence, Section 4 supplies a cost model the buyer fills in, Section 5 sets out four named programmes, and Sections 6 to 8 convert the gaps into risks, a platform comparison and a recommended posture. Every figure that originates with a manufacturer, OEM or vendor carries a company-claimed label. Where a figure could not be found, the report says so rather than estimating it.


Five signals buyers should act on

1. Evidence measures activity, not economics. The BMW totals (30,000+ X3 vehicles supported, 90,000+ components moved, about 1,250 hours) come from BMW and Figure AI only (company-claimed). None of these figures carries a cost baseline, so none shows a saving. They show that a task was performed, not that it paid.


2. The three programmes are not equivalent. BMW completed a company-reported Figure 02 pilot. Mercedes-Benz is in test mode with "fewer than ten" Apptronik Apollo units (company-claimed relay). Foxconn's Houston pilot, also "fewer than 10" humanoids, was planned for mid-April 2026; its start and results are not found. A league table of the three would mislead.


3. Capital runs ahead of evidence. Figure AI's $39B valuation (September 2025, company-announced), Apptronik's $935M Series A total and Agility Robotics' $2.5B SPAC value (special-purpose acquisition company) sit against no independently verified outcome data. Funding shows access to capital, not customer-level economics.


4. Commercial terms are opaque. No price, lease or RaaS (robots-as-a-service) term was found for any Western platform. Buyers must define the metrics and the cost model themselves.


5. Volume evidence sits with Chinese vendors. Unitree shipped more than 5,500 humanoids in 2025 and UBTech sold 1,079 full-size Walker S units in its FY2025 results (both verified, reported secondhand from the prospectus and results). Western OEM evidence remains single-site and small-unit.


2. Platform / Market Landscape

Annual humanoid shipments are below 100,000 units, according to Interact Analysis (21 May 2026; a published estimate, methodology not disclosed). About 10% of 2025 production was deployed in real-world operations. A secondary report puts H1 2026 shipments at about 19,100 (unverified; definition not given). Forecast series differ in period and definition, so they should be read side by side and never averaged.


Market-level evidence is estimate-based and thin on definitions. Interact Analysis (21 May 2026) forecasts a commercial inflection in 2032, with more than 700,000 annual units and about $15B revenue by 2035, conditional on economic-viability thresholds, embodied-AI progress, clearer regulation and acceptable efficiency rates (published forecast; methodology not disclosed). McKinsey (15 October 2025) counted fewer than ten of about 50 companies with humanoid ambitions at scaled pilots or precommercial deployment (analyst estimate). Gartner, relayed by The Robot Report (28 January 2026), expects fewer than 100 companies to move humanoid proofs of concept beyond experimentation through 2028, from about 200 proofs of concept, and fewer than 20 to have humanoids live in production in supply chain or manufacturing (analyst forecast).


The International Federation of Robotics (IFR) frames humanoids as a promising technology where flexibility is required (Business Wire, 8 January 2026). Its position paper expects humanoids to complement existing robot types rather than replace them, with mass adoption that "remains uncertain" (14 August 2025; both verified as IFR statements). IFR's 2025 all-robot figures (operational stock of 5 million, up 9%; more than 600,000 installed, up 11%) are not humanoid figures and should not be read as such.

Source and date

Period

Figure

Label

Interact Analysis, 21 May 2026

Annual, current

Below 100,000 units; about 10% of 2025 production deployed in real-world operations

Verified published estimate; methodology not disclosed

TechNode, 11 Aug 2026 (cites IT Home)

H1 2026

About 19,100 units (vs about 5,100 in H1 2025); Chinese makers above 97%

Unverified secondary; "shipments" undefined

Counterpoint Research, 21 Aug 2026

2026 to date

"Up 300%"; "up to 30,000"

Headline only; no baseline or method (company-claimed)

Goldman Sachs via 24/7 Wall St, 14 Sep 2026

2026 forecast

75,000 units (prior 51,000); 890,000 in 2030; about 6.5M in 2035

Secondhand analyst forecast (company-claimed)

TechNode, 11 Aug 2026

FY2026 forecast

About 60,000; about 500,000 by 2030

No forecaster named (company-claimed)

 

The series cannot be reconciled because period, definition and method differ. Do not average them. Verified volume comes from filings, reported secondhand: Unitree's prospectus shows more than 5,500 humanoids shipped in 2025 (excluding wheeled dual-arm models), RMB 1.70B revenue and RMB 278.21M net profit; UBTech's results show 1,079 full-size units sold, RMB 821M of full-size humanoid revenue (2024: RMB 35.6M) and annual capacity above 6,000 by end-2025. Shipment counts measure units sold, not task outcomes, so none of these figures supports a factory-level return claim.


Three readings follow for a buyer.

  • First, the volume leaders are not the vendors whose factory programmes dominate Western coverage. Chinese makers accounted for more than 97% of H1 2026 shipments in the one secondary source that gives a split (unverified), so a buyer comparing Western pilots with global volume is comparing different markets.

  • Second, McKinsey characterises Unitree pricing as under $10,000 (analyst characterisation), which indicates that unit price and task capability are separate questions.

  • Third, the gap between shipments and deployment matters most: Interact Analysis puts real-world deployment at about 10% of 2025 production, so most units sold were not working in operations. A buyer should ask any vendor quoting shipments how many units are in paid, task-bearing use.

Vendor

Capital event

Outcome data disclosed

Figure AI

Series C above $1B at $39B (16 Sep 2025, company-announced); reported talks of about $1B more from NVIDIA (7 Oct 2026, unconfirmed)

BMW totals, company-reported only

Apptronik

$520M extension; Series A total $935M (February 2026); valuation reported at $5.3B–$5.5B+, not disclosed by Apptronik

None

Agility Robotics

SPAC with Churchill Capital Corp XI at $2.5B; more than $620M expected cash, including about $200M PIPE led by Foxconn (24 Jun 2026)

More than 65,000 cumulative operating hours (company-claimed); no outcome metrics

Humanoid (UK)

$152M Series A at $1.35B (22 Jul 2026)

Siemens pilot figures are the companies' own

 Agility's filing states that its more than $300M of booked orders are not a measure of current-period revenue and depend on milestones (company-claimed). Production targets are targets, not shipments: Figure's initial BotQ factory target is 12,000 robots a year (company-claimed), and Hyundai cites capacity of 30,000 robots a year (company-claimed). The [internal link: Quarterly Humanoid Platform Index] holds platform-level specifications; this report does not duplicate them.


3. Deployment Evidence: Reported Outcomes, Not Just Announcements

No named manufacturing deployment has independently verified outcome data. BMW and Figure AI report more than 30,000 X3 vehicles supported, more than 90,000 components moved and about 1,250 operating hours (company-claimed). Achieved success rate, uptime and intervention rate are not found. Mercedes-Benz and Foxconn have announcements and small unit counts without outcomes. A buyer should place every programme on an evidence ladder before using it as a reference.


A deployment announcement, a pilot and a sustained operation are different evidence types. The ladder below separates them, and the placements are this report's own classification, not vendor statements.

• Rung 1 — Planned: Announced; not confirmed started. Placed here: Figure 03 (BMW), AEON (BMW), Foxconn Houston

• Rung 2 — Demonstrated: Task shown; test mode. Placed here: Apollo (Mercedes-Benz)

• Rung 3 — Piloted: Run in a plant for a defined period; totals reported. Placed here: Figure 02 (BMW), company-reported

• Rung 4 — Sustained operation: Continued use with disclosed operating data. Placed here: None found

• Rung 5 — Independently verified: Outcome measured or audited by a neutral party. Placed here: None found


Two features of the ladder deserve emphasis. No programme in the record reaches rung 4 or rung 5: none discloses operating data over a sustained period, and none has had its outcomes measured by a neutral party. The highest placement, Figure 02 at rung 3, also ended. Figure retired the fleet after eleven months, so the only piloted programme is no longer running in its original form, and its successor sits at rung 1. Figure 03 is therefore a planned deployment, not a reference site.


What each programme has reported, and what a buyer needs

 

Attribution discipline

"Supported" is not "built". The 30,000+ X3 figure describes vehicles produced at a plant during a deployment in which a BMW spokesperson said, as of April 2025, that one robot worked at a time. The wording the report applies, and the reason, is below.

Claim

Origin

Status

Required wording

30,000+ X3 vehicles

BMW and Figure AI

Company-claimed; independent verification not available per The Robot Report (29 Jun 2026)

"Vehicles supported during the deployment"

90,000+ components moved

BMW and Figure AI

Company-claimed

Define the task as sheet-metal loading

About 1,250 operating hours

BMW and Figure AI

Company-claimed; 10 months (BMW) or 11 (Figure)

Distinguish runtime from scheduled hours and productive uptime

84-second cycle, 37 seconds for loading; 5 mm tolerance; ">99%" accuracy

Figure, via Humanoids Daily (19 Nov 2025)

Design figures and a target; achieved rate not reported

"Design figure" or "target", never "result"

Autonomy as the goal

Mercedes-Benz

Units are teleoperated; autonomy is a stated objective

Do not describe routine production autonomy

Deployment "within six months or so"

Foxconn chairman Young Liu, via The Register (7 Nov 2025)

Company-claimed; Nikkei original not fetched

Attribute; do not call it a missed deadline

 

Denominator check (this report's arithmetic, to be put to the vendor; not a finding). Figure's 84-second cycle implies about 43 cycles an hour, or about 54,000 cycles over 1,250 hours. The reported 90,000+ components cannot be reconciled with that without knowing the number of parts per cycle, which is not disclosed. Separately, 1,250 operating hours over about ten months compares with roughly 2,150 scheduled shift hours (10 hours × 5 days × about 43 weeks). The gap is unexplained in the sources.


The practical test for every claim is the denominator: how many robots, which task, over what scheduled period, with what level of human assistance.

Four questions turn that test into a vendor conversation.

  • How many robots produced the total, and were they working at the same time?

  • What share of scheduled hours was the robot actually productive, as distinct from powered on?

  • What did a person do during the run: loading, resetting, remote takeover or supervision?

  • And what did the same task cost before the robot arrived?


A vendor that answers the first two but not the last two has described activity. A vendor that answers all four has started to describe economics.


Vendor-reported performance figures from outside the three programmes follow the same pattern. Humanoid (UK) and Siemens report a two-week Erlangen pilot with 60 tote moves an hour, more than 90% autonomous pick-and-place success (exact figure not given), more than 8 hours of uptime and more than 30 minutes of continuous autonomous execution (DC Velocity, 15 Jan 2026; company-claimed). AgiBot reports a six-day livestream at Longcheer with 64,828 tasks and 99.99% success (eWeek, 9 Jul 2026; company-claimed), and Xiaomi reports 98% task success on an automotive line without a sample size, duration or method (EqualOcean, 15 Jul 2026; company-claimed). These success rates lack denominators, task definitions and error-handling rules, so they cannot be compared across vendors or read against the BMW totals.


The Journal's earlier deployment evidence report (www.physicalaijournal.org/post/humanoid-deployment-evidence-report-2026) covers a wider platform set. The evidence standard follows the https://www.physicalaijournal.org/editorial-standards page.


4. Economics & ROI Analysis

Cost per productive hour equals total annual cost (acquisition or lease, integration, maintenance, supervision, tooling, energy and downtime) divided by productive hours, compared with the best conventional alternative. No verified price, lease term or uptime figure exists for any Western platform, so every input is buyer-entered. Throughput totals such as BMW's are activity counts, not return on investment (ROI), and a model is only as sound as its inputs.


4.1 Why throughput is not ROI

BMW's reported totals describe what a robot handled, not what it cost or what it displaced. A cost baseline needs acquisition or lease cost, integration engineering, supervision, maintenance, downtime and the labour redeployed. None is disclosed for any programme in this report.

A simple price comparison is misleading. McKinsey puts prototype cost at about $150,000–$500,000 and the price at which a humanoid competes with human labour at about $20,000–$50,000, and it attributes 40–60% of cost to actuation (analyst estimates). Mercedes-Benz's Jörg Burzer made the same point from the buyer side (Reuters, 18 March 2025; reported fragment). Neither statement is a price quote. Both indicate that today's economics are a function of price decline that has not yet been demonstrated, so a buyer modelling a pilot on a future price is modelling a forecast.


4.2 The cost stack and what the record shows

Variable

Status

What the public record shows

Acquisition or lease

Public anchor / not found

McKinsey (15 Oct 2025): prototype cost about $150,000–$500,000; price needed to compete with human labour about $20,000–$50,000 (analyst estimates). List, lease and RaaS terms: not found for Figure, Agility, Apptronik, Atlas, Hexagon or Foxconn

Integration and workcell engineering

Buyer-entered

Not found. Foxconn's engineer described fixed grippers now, with end effectors swapped per task

Maintenance and spares

Buyer-entered

Not found

Supervision or remote-operation labour

Buyer-entered

Not found; Mercedes-Benz units are teleoperated in the reports retrieved

Energy and charging

Partial

Company-claimed battery runtime of about 4 hours for Apollo and AEON; McKinsey says most humanoids run 2–4 hours per charge against 8–12-hour shifts

Scheduled hours

Partial

10-hour Monday–Friday shifts at BMW Spartanburg (company-claimed)

Productive availability

Buyer-entered

No verified uptime for any platform

First-pass quality

Buyer-entered

Not found; ">99%" is a Figure target

Downtime recovery

Buyer-entered

Not found

 

Other anchors are limited. Secondary reports put Mercedes-Benz's Apollo unit price in the tens of thousands of euros (heise; Xpert.digital, 2025), treated as unverified. Mercedes-Benz production chief Jörg Burzer reportedly linked interest to unit cost falling to a two-digit-thousand-dollar level, a condition, not a commitment (Reuters via AOL, 18 March 2025; reported fragment) . Gartner states that humanoids cost multiples of task-specific robots with lower throughput and uptime (analyst forecast; no dollar figures).


Battery runtime is the least visible cost driver. Apollo and AEON both carry company-claimed batteries of about 4 hours, and McKinsey says most humanoids run 2–4 hours per charge against 8–12-hour shifts (analyst estimate). A 10-hour shift therefore implies at least two swaps or charging stops. Hexagon claims a swap takes about 26 seconds, and Boston Dynamics describes a self-swapping Atlas battery (company-claimed). Each swap either costs productive time or requires a person or a spare unit, and none of the programmes in this report discloses which. Supervision is the second hidden driver: a unit that is teleoperated, as the Mercedes-Benz Apollo units are reported to be, adds robot cost to a labour cost that remains.


4.3 Formulas

Cost per productive hour = (A + I + M + S + T + E + D) ÷ H

A = annualised acquisition or lease · I = integration and workcell engineering, annualised · M = maintenance and spares · S = supervision and remote-operation labour · T = end effectors and tooling · E = energy and charging · D = downtime recovery · H = productive hours = scheduled hours × productive availability

Cost per accepted task = total annual cost ÷ (attempted tasks × first-pass quality)
Break-even productive hours = fixed annual cost ÷ (benchmark cost per productive hour of the best conventional alternative − variable cost per productive hour)

4.4 Sensitivity: how fragile the answer is

The grid uses illustrative input, not market data. It assumes 4,000 scheduled hours a year, treats all cost as fixed, and tests three annual cost cases against three availability levels. Replace every value with vendor-committed figures.

Cost per productive hour (illustrative)

50% availability (2,000 h)

70% availability (2,800 h)

90% availability (3,600 h)

Conservative: $150,000 a year

$75.00

$53.57

$41.67

Base: $100,000 a year

$50.00

$35.71

$27.78

Optimistic: $60,000 a year

$30.00

$21.43

$16.67

 

Against an illustrative benchmark of $30 per productive hour, the base case breaks even only above about 83% availability, the optimistic case at 50%, and the conservative case never does, because its cost per hour is $37.50 even at full availability. The point is sensitivity, not a forecast: availability is the least documented variable and moves the result most.

Read the grid in rows and in columns. Across a row, moving availability from 50% to 90% cuts the cost per productive hour by about 44% in every scenario, because cost is held fixed. Down a column, moving from the conservative to the optimistic case cuts it by 60%. Availability therefore moves the answer about as much as the cost of the robot does, yet vendors publish prices more readily than availability. A buyer comparing two quotes should ask for the availability behind each before comparing the prices. The inputs here are placeholders chosen to show the arithmetic, and no vendor, programme or market figure sits behind them.


4.5 What a vendor must disclose before the model runs

Five written disclosures make the model usable: price or lease terms and conversion rights; integration scope and engineering hours; measured availability, intervention and first-pass quality data from a comparable site, with the denominator; the maintenance and spares plan; and the autonomy level, including any teleoperation. A vendor that cannot supply these cannot be modelled, and the buyer should treat the pilot as evidence-gathering rather than investment.


5. Named Deployment Case Studies

BMW completed a company-reported Figure 02 pilot and announced Figure 03 and Hexagon AEON follow-ons. Mercedes-Benz tests fewer than ten Apptronik Apollo units, with no outcomes disclosed. Foxconn planned a Houston pilot of fewer than 10 humanoids, with start and results not found. Only BMW has quantified totals, and all are company-reported. No programme has independently verified outcomes, and none discloses uptime, success rate or intervention rate.


5.1 BMW and Figure AI, Plant Spartanburg, South Carolina

BMW and Figure AI announced a commercial agreement on 23 January 2024, described by Fortune (6 April 2025) as milestone-based, with financial terms and duration undisclosed. Figure 02 inserted sheet-metal parts for welding in the body shop on 10-hour Monday–Friday shifts. BMW says the deployment ran about ten months in 2025; Figure says eleven (task, shift pattern and durations are company-claimed). Until about March 2025 the robot worked only in non-production hours, and a BMW spokesperson, Steve Wilson, told Fortune that one robot worked at a time. Fortune also reported that Figure chief executive Brett Adcock's references to a "fleet" went beyond BMW's account; Humanoids Daily reports that Adcock disputed this.


All three totals trace to BMW and Figure only (company-claimed). Figure retired the Figure 02 fleet after eleven months (about November 2025; Figure-reported, company-claimed). On 25 June 2026 BMW announced that Figure 03 would handle logistics sequencing, picking unsorted parts from containers and sorting them onto a trolley for the line. BMW's release gives no robot count, go-live date, terms or metrics, and live status is not found.

BMW Manufacturing's Ulrich Wieland, Vice President, Production Control and Logistics, said:

"Plant Spartanburg is the birthplace of humanoid robotics in BMW Manufacturing's operational day-to-day activities."

(BMW PressClub, 25 June 2026; company characterisation)


Ladder stage: Figure 02, piloted; Figure 03, planned. A buyer can conclude that a humanoid performed a bounded loading task in a live plant. A buyer cannot conclude productivity gain, cost saving, achieved success rate or fleet scalability.


5.2 BMW and Hexagon Robotics AEON, Leipzig, Germany

Hexagon Robotics announced on 27 February 2026 that its wheeled humanoid AEON would be BMW's first humanoid pilot in German production, for high-voltage battery assembly and exterior component manufacturing (company-claimed). Company statements describe a first test in December 2025, a further test in April and a pilot phase in "summer" 2026. The unit count is undisclosed; an AFP summary says two, which is unconfirmed. BMW describes a staged evaluation: laboratory testing, plant trials, then pilots, with an aim of pilots at other plants (company-claimed). Whether the summer pilot started is not found, and a trade headline's reference to "series production" is not supported by its own text.

Ladder stage: planned. A buyer can conclude that BMW is testing a second, wheeled form factor. A buyer cannot conclude performance, unit count or pilot status.


5.3 Mercedes-Benz and Apptronik, Berlin-Marienfelde and Kecskemét

Mercedes-Benz and Apptronik announced a commercial pilot agreement on 15 March 2024, without terms. On 18 March 2025 Reuters reported an equity investment, described as a "low double-digit million-euro sum" (exact amount and ownership not found), and testing at Berlin-Marienfelde, Germany, and Kecskemét, Hungary. Apollo units move parts to the line and run preliminary quality checks (company-claimed). heise reports fewer than ten units (relay), trained through teleoperation with augmented reality, with autonomy the goal. Burzer told journalists in Berlin:

"I'd like to see us find applications in areas where we face labour shortages"

(Reuters, 18 March 2025; historical) .


Mercedes-Benz took part in Apptronik's $520M extension (February 2026) as an existing investor; its share is undisclosed. No outcome data, pilot length or conversion to purchase is found, and the primary Mercedes-Benz sources are from 2024 and 2025. A 9 October 2026 Autoplant Brief report of live assembly-line deployment lacks corroboration and discloses no scope, count, uptime or throughput.

Ladder stage: demonstrated (test mode). A buyer can conclude that a tier-one OEM announced an Apollo pilot in March 2024 and invested in the vendor. A buyer cannot conclude that Mercedes-Benz has validated Apollo's economics.


5.4 Foxconn and NVIDIA, Houston, Texas

On 28 October 2025 Foxconn said its Houston plant, which makes AI servers for NVIDIA, would be among the first to deploy humanoids powered by NVIDIA's Isaac GR00T N model (Reuters; company-claimed). A Foxconn engineer told Instrumental (20 April 2026) that the pilot would involve fewer than 10 humanoids on GB300 server assembly, with part transfer, screw fastening and materials handling demonstrated, fixed grippers for now and training in NVIDIA Isaac Sim, and that no timeline exists for harder tasks. Chairman Young Liu told Nikkei in November 2025 that deployment would come within about six months. Reuters had reported anonymous sources in June 2025 targeting Q1 2026; the shift to mid-April 2026 is an inference, not a documented slip.


Whether the pilot started, its results and the humanoid hardware model are not found. Reuters coverage of Foxconn's Q2 2026 profit (12 August 2026) does not mention humanoids; that silence is not evidence. A report of a Foxconn humanoid showing at NVIDIA GTC 2026 could not be verified from the sources retrieved. Foxconn also led a roughly $200M private placement in Agility Robotics' SPAC (June 2026), which signals capital involvement, not deployment outcomes.


Ladder stage: planned. A buyer can conclude that Foxconn is scoping a small, task-specific pilot. A buyer cannot conclude that any Foxconn humanoid runs in production.


Timeline

Date

Event

Ladder stage

Label

23 Jan 2024

BMW–Figure commercial agreement announced

Planned

Terms undisclosed (verified)

15 Mar 2024

Mercedes-Benz–Apptronik pilot agreement

Planned

Company-claimed

18 Mar 2025

Mercedes-Benz stake in Apptronik; testing at two plants

Demonstrated

Reuters (statement made)

2025

Figure 02 pilot, Spartanburg (10–11 months)

Piloted

Company-claimed

28 Oct 2025

Foxconn Houston deployment announced

Planned

Company-claimed

About Nov 2025

Figure retires Figure 02 fleet

Piloted (ended)

Figure-reported

Dec 2025

First AEON test, Leipzig

Planned

Company-claimed

27 Feb 2026

AEON pilot announced

Planned

Company-claimed

Mid-Apr 2026

Foxconn Houston pilot planned; start not found

Planned

Company-claimed

25 Jun 2026

Figure 03 sequencing announced; live status not found

Planned

Company-claimed

"Summer" 2026

AEON pilot phase; start not found

Planned

Company-claimed

6. Friction, Risk & Unresolved Issues

The main risks are evidentiary and commercial. No independent outcome verification exists, achieved reliability metrics are not published, commercial terms are opaque, and several programmes rely on teleoperation or single-cell tasks. Reported timelines have slipped at Tesla and Apptronik (secondary sources). Gartner expects fewer than 20 companies to have humanoids live in supply-chain or manufacturing production through 2028 (analyst forecast). A buyer should treat each risk as a pilot design input.


Risks reach a buyer in the order they would surface during a pilot: scope, supervision, schedule, credibility, labour relations and standards. The register below pairs each with the evidence and a mitigation the buyer controls.

Risk

Evidence

Label

Buyer mitigation

No independent outcome verification

BMW totals trace to BMW and Figure; Mercedes-Benz and Foxconn have no outcome data

Company-claimed; none verified

Make independent measurement a pilot term

Reliability metrics absent

Success rate, uptime and intervention rate not found for Figure; no uptime data from Apptronik or Mercedes-Benz

Company-claimed

Write metric definitions into the contract before the pilot

Teleoperation dependence

Mercedes-Benz Apollo units reported as teleoperated; ZeroGantry describes "data-collection mode, not unsupervised production"

Company-claimed

Require the autonomy level to be stated per task

Single-cell scope

Figure 02 worked a contained cell in an already automated body shop; 84-second and 37-second figures describe one task, not line throughput

Verified (BMW statement made); content company-claimed

Do not extrapolate to line productivity

Timeline slippage

Tesla: "late July or August" (22 Apr 2026), then anticipation of production in 2026 (22 Jul 2026, summary wording), then "production began in Fremont" (Q3 deck, 21 Oct 2026, relayed, deck not opened); Apptronik missed an end-2025 successor target; Foxconn moved from Q1 2026 (anonymous) to mid-April 2026 (inference)

Company-claimed; secondary-only

Tie payment to milestones, not announcements

Opaque commercial terms

No price, lease, service-level or warranty term for BMW, Mercedes-Benz or Foxconn; Agility's more than $300M of orders (company-claimed) depend on milestones

Verified as undisclosed

Request terms in writing before a pilot

Vendor-claim dispute

Fortune (6 Apr 2025) said Adcock's "fleet" and "running everyday" claims went beyond BMW's account; Humanoids Daily says Adcock disputed this

Verified (statements made)

Use the operator's account, not the vendor's, as the baseline

Labour relations

The Hyundai Motor branch of the Korean Metal Workers' Union said "Not a single robot can enter the production floor" without a labour-management agreement (22 Jan 2026, before the report window; Hyundai's reply not reported)

Company-claimed (union statement)

Engage employee representatives before a site trial

Standards status

ISO/CD 25785-1 (ISO/TC 299) sits at Committee Draft stage (30.20) per ISO.org, with no date on the page; Part 2 (integration) is yet to be developed; no ANSI/A3 humanoid standard found

Stage verified per ISO page; timelines secondary

Report only; qualified engineers assess the site

 

Three of these risks deserve a longer note because they change how a pilot is designed.

  • The first is supervision. A unit that needs a person to assist or take over is still a working humanoid, but its economics are different, and a vendor total that does not count interventions hides the difference. The buyer should log every assist from the first shift, whatever the vendor's definition.

  • The second is scope. Figure 02 worked inside a contained cell in a body shop that was already highly automated, and the task was a single loading step. A result from that setting says little about a mixed-task workstation or a line with people moving through it.

  • The third is schedule. Tesla's sequence of statements moved from a stated production window to a stated anticipation to a reported start, and the Q3 deck that reports the start was not opened and gives no rate, unit count or price. A buyer who schedules around a vendor timeline without a milestone-linked payment carries the slippage.


The credibility dispute is a different kind of risk. When the operator's account and the vendor's account differ, as Fortune reported in April 2025, the buyer needs a rule for which prevails. The report's rule is that the operator's statement sets the baseline and the vendor's statement is treated as a claim to be tested. That rule applies equally to every vendor named here.


The standards row reports status and nothing else. This report offers no view on the conformity of any platform, and safety sign-off remains a gate for the buyer's own qualified engineers.

Analyst scepticism is consistent. Gartner's Abdil Tunca, Senior Principal Analyst, said:

"The promise of humanoid robots is compelling, but the reality is that the technology remains immature"

Abdil Tunca, Gartner Supply Chain practice, via The Robot Report, 28 Jan 2026 (historical; analyst view)

Gartner's forecast sets the scale of the gap. From about 200 proofs of concept, it expects fewer than 100 companies to move beyond experimentation through 2028, and fewer than 20 to have humanoids live in supply-chain or manufacturing production (analyst forecast, relayed by The Robot Report).


A buyer entering a pilot in 2026–2027 is therefore joining a group in which most participants, on this forecast, will not reach production. Goldman Sachs lists reliability, training data, autonomy and cost as obstacles (secondhand, 14 Sep 2026). Labour-displacement claims need a sector, a timeframe and a confidence tier; no source retrieved supports one for manufacturing.


Source conflicts that affect a buyer

Item

Conflict

Handling in this report

Figure 02 robot count

One at a time (BMW's Steve Wilson, Apr 2025); a single unit (The Robot Report, 2 Mar 2026); "two" (IndustrialSage, no named source); "fleet" (Adcock)

Unresolved; BMW's account cited as of April 2025 only

Pilot year and length

2024 (The Robot Report) against 2025 (BMW); ten months (BMW) against eleven (Figure)

BMW's year used; both durations attributed

Figure 03 status

"Planned" (BMW) against "deploys" (Automotive World; Robot Report headline)

BMW's wording used

Apptronik valuation

About $5B to more than $5.5B across outlets; Apptronik disclosed none

Cited as $5.3B–$5.5B+, reported

Mercedes-Benz status

Teleoperated test against "live production" (Autoplant Brief, 9 Oct 2026)

Former used; latter unconfirmed

Shipments

19,100 (H1), "up to 30,000", below 100,000, 75,000 forecast

Shown side by side, never averaged

 

7. Competitive / Platform Comparison

Task variability, not general claims of flexibility, decides when a humanoid robot should replace a fixed industrial robot, cobot or AMR. Humanoids fit variable tasks in a space built for people, where a conventional alternative cannot meet the requirement. The International Federation of Robotics (IFR) describes humanoids as complementing, not replacing, existing robots. Otherwise a conventional option usually wins on cost and throughput, according to Gartner (analyst forecast).


7.1 Readiness matrix

The matrix rates how well each programme's public record documents six dimensions. Ratings are Documented, Partial or Not disclosed; they are not performance scores, and no numeric score is possible because no comparable measure exists in the record. Foxconn appears as a programme row because its hardware model is not identified.

Programme

Task defined

Units in use

Operating data

Reliability metrics

Commercial terms

Independent verification

BMW / Figure AI (Figure 02)

Documented

Partial

Partial

Not disclosed

Not disclosed

Not disclosed

BMW / Hexagon AEON

Documented

Not disclosed

Not disclosed

Not disclosed

Not disclosed

Not disclosed

Mercedes-Benz / Apptronik Apollo

Partial

Partial

Not disclosed

Not disclosed

Not disclosed

Not disclosed

Foxconn Houston (programme only)

Partial

Partial

Not disclosed

Not disclosed

Not disclosed

Not disclosed

 

Read the matrix by column. The first column is the only one where any programme is fully documented, which means the public record describes what each robot was meant to do far better than what it did. The commercial-terms and independent-verification columns are empty for every programme, and those two columns decide whether a purchase case can be written. A buyer should treat each "Not disclosed" cell as a question for the vendor, to be answered in writing before a pilot starts.


The full platform set sits in the [internal link: Quarterly Humanoid Platform Index]. Other platforms in the record, including Agility's Digit, Boston Dynamics' production Atlas, Tesla's Optimus and the UBTech Walker S series, have their own company-claimed specifications and none has a disclosed price. Boston Dynamics states that its 2026 Atlas output is committed to Hyundai's RMAC and Google DeepMind, with further customers from early 2027 (company-claimed).


7.2 Manufacturer-claimed specifications

Platform

Manufacturer-claimed data

Fields not found

Figure 03

Soft components, wireless charging, speech-to-speech audio, improved hands with tactile sensors and palm cameras (per Adcock; The Robot Report, 29 Jun 2026)

Payload, runtime, units delivered, price

Apptronik Apollo

Company page: 160 lb and 55 lb (labels ambiguous), about 4-hour battery, hot-swappable packs. Secondary (Xpert.digital, 21 Mar 2025): 1.73 m, 72 kg, 25 kg payload, teleoperated

Height and hands (company page), price

Hexagon AEON

1.65 m, 60 kg, top speed 2.4 m/s (bmwgroup.com says about 2.5 m/s), payload 15 kg short-term and 8 kg long-term, 4-hour battery, about 26-second battery swap

Price, unit count

Foxconn Houston humanoid

Hardware model not identified

Everything

 

All specifications are company-claimed. This report makes no engineering or safety assessment of any listed platform. Two conflicts in the table matter to a buyer. Apptronik's own page and a secondary source give different weights and payloads for Apollo, and the labels on the company page are ambiguous, so a spec sheet should be requested for the unit actually offered. BMW and Hexagon give different top speeds for AEON (2.4 m/s and about 2.5 m/s). The difference is small, but it shows that even a primary-source spec needs a dated, versioned data sheet before it is used in a requirement.


7.3 Humanoid or conventional automation

The table is an editorial framework for task-fit screening. Its ratings are qualitative and not measured values.

Option

Task variability

Cycle time

Integration need

Cost basis

Humanoid robot

Suited to high variability in people-built space

Not documented against a baseline

Not documented; fixed grippers and per-task end effectors reported at Foxconn

No verified price; analyst ranges only

Fixed industrial robot

Suited to low variability, repeatable tasks

Not documented in this record

Not documented in this record

Not in this record

Cobot

Moderate variability in shared space

Not documented in this record

Not documented in this record

Not in this record

AMR

Moves material between points; not a manipulator

Not documented in this record

Not documented in this record

Not in this record

Purpose-built handling

Fixed task

Not documented in this record

Not documented in this record

Not in this record

 

Given Gartner's cost-multiple statement (analyst forecast), a humanoid case has to rest on a requirement the conventional option cannot meet, not on a general preference for flexibility. Where a fixed robot, cobot or AMR meets the requirement, the burden of proof sits with the humanoid.


The screening logic is a sequence of questions.

Is the task stable enough to be built into a fixture or a line? If so, a fixed fixture or line is the first alternative to price.

Does the task move material between fixed points? If so, an AMR or a conveyor addresses it.

Does it need a manipulator in a space shared with people and already laid out for them, where changing the layout would cost more than the robot? Only then does a humanoid enter the comparison, and even there the question is whether a cobot on a mobile base does the job.

IFR's position, that humanoids are a promising technology where flexibility is required and complement existing robot types, is consistent with this order.


The evidence also limits how far a buyer can compare. A Foxconn engineer described tooling that changes per task (company-claimed, one site), so a buyer should ask each vendor which tooling a given task requires.


8. Strategic Recommendations (Automation/Procurement Buyer)

For most buyers, run a bounded paid pilot on a narrow, repeatable task with defined acceptance metrics and exit terms, and defer production-scale capital. The public record contains no independently verified economics at any named site. A purchase case requires disclosed terms and measured results. This posture is conditional, applies to manufacturing in 2026–2027, and carries a medium confidence tier because it rests on absence of evidence.


8.1 The thesis

The recommended posture is conditional: fund bounded paid pilots for narrow, repeatable tasks, and defer production-scale capital until independently measured results and disclosed terms exist. The record shows that humanoids can perform defined tasks. It does not show what those tasks cost, what they displaced or how often a person had to step in. Capital committed before those answers exist buys the vendor's learning, not the buyer's return. This conclusion is the report's own and is subject to fact check; it is not a labour-displacement forecast.


A paid pilot differs from a demonstration in three ways. The buyer sets the task and the metrics, the buyer's own people measure the result, and the contract states what happens at the end. A vendor-run demonstration, however polished, answers the vendor's question. The pilot should be sized to produce evidence: one task, one cell, a fixed number of scheduled shifts, and a comparison against the current method run over the same period. Foxconn's reported approach, fewer than 10 units on a single server-assembly task with no timeline for harder tasks, is a reasonable scale for a first test, though its outcome is not yet known.


The posture is conditional on a medium confidence tier for manufacturing in 2026–2027. It would weaken if a vendor supplied independently audited operating data at a comparable site, and it would strengthen if announced pilots stayed unconfirmed.


8.2 Buy, pilot or wait


8.3 Pilot acceptance scorecard

Every threshold is buyer-set, because no published pilot-acceptance standard exists. McKinsey's indicator list (certified fenceless safety, shift-equivalent uptime at customer sites, dexterity benchmarks, bill-of-materials cost curve, pilot-to-scale conversion rates, binding orders against letters of intent) is the only public anchor (analyst view).

Metric

Definition

Buyer-set target

Pass / extend / terminate

Productive availability

Productive hours ÷ scheduled hours

Set before pilot

Pass at or above target; extend within a stated band; terminate below

Successful cycles per shift

Accepted cycles, counted by the buyer

Set before pilot

As above

First-pass quality

Accepted tasks ÷ attempted tasks

Set before pilot

As above

Human interventions per shift

Every assist, reset or remote takeover

Set before pilot

Pass at or below target

Recovery time

Time from fault to productive restart

Set before pilot

Pass at or below target

Integration cost

Actual against quoted

Quoted figure

Terminate on a stated overrun

Payback

Total cost against benchmark cost per productive hour

Set before pilot

Computed with Section 4 formulas

 

Use the scorecard as a decision rule, not a report. Fix every target before the first shift, because targets set after seeing results drift toward whatever the pilot achieved. A result that passes on cycles but fails on interventions should extend the pilot, not scale it, since a high cycle count with frequent assists describes a supervised process. A result that fails on integration cost should terminate, because integration is the cost the vendor is least likely to have quoted accurately. Record failures with a cause: hardware, software, task definition or site. Without a cause, the buyer cannot tell whether a failed pilot argues against the robot or against the task.


8.4 Vendor evidence request list

Seven evidence items should be in hand before any paid pilot: operating data from a comparable site with robot count, task, scheduled hours and intervention log; availability and first-pass quality with definitions; the autonomy level and any teleoperation; the maintenance and spares plan; integration scope and engineering hours; price, lease or service terms; and customer references the buyer may call. McKinsey's procurement screens add binding purchase orders against letters of intent and pilot-to-scale conversion rates, which test whether a vendor's other customers have committed money or only interest (analyst view). A vendor that cannot or will not supply the first group has not met the evidence standard.


8.5 Contract-protection checklist

This is a procurement aid for legal review, not legal advice. Items: performance definitions and buyer data access; availability and repair-time commitments; spares; software-update and change-control terms; conversion terms from pilot to purchase; exit rights; ownership of operating and training data.

Two clauses matter more than the rest. The first is the performance definition: terms such as "uptime" or "success" must be defined in the contract, with who measures them, because the public record shows that undefined terms produce totals nobody can reconcile. The second is the exit: if the pilot ends without conversion, the buyer should keep its operating data, retain access to anything it paid for, and owe nothing beyond the pilot fee. Counsel should draft the language, and the safety and integration sign-off belongs with the buyer's qualified engineers, not the commercial team.


8.6 Triggers that would change the posture

Four developments would move the posture toward scale: an independent audit of operating data; published intervention and uptime data; disclosed price or lease terms; or movement of ISO/CD 25785-1 beyond Committee Draft (reported only). Move toward further delay if announced pilots stay unconfirmed or terms stay undisclosed.

Next action: define the pilot scorecard and request the vendor evidence pack before any capital approval.


9. Executive FAQ

A bounded paid pilot, not production-scale capital, is the supported decision for most buyers of humanoid robots in 2026–2027. The six answers below cover cost per hour, return on investment, the BMW deployment, conventional alternatives, pilot acceptance criteria and timing. The evidence base closes on 23 October 2026. Each keeps company-claimed labels, states missing figures as not found, and implies no independent verification where none exists.


How much does a humanoid robot cost per hour in a factory?

No verified cost per hour exists. List and lease prices are not found for any Western platform, and no operator has published availability or intervention data against which to divide a cost. McKinsey's ranges (prototype about $150,000–$500,000; price to compete about $20,000–$50,000) are analyst figures. Use the Section 4 formula and enter your own inputs.


Is there evidence of ROI from humanoid robots at BMW, Mercedes-Benz or Foxconn?

No. BMW reports activity totals (company-claimed): more than 30,000 X3 vehicles supported and about 1,250 operating hours. Mercedes-Benz and Foxconn report none. No cost baseline or independent verification is published for any of the three, so none supports a return claim, and a buyer should not cite them as one.


What did Figure 02 actually do at BMW Spartanburg?

Figure 02 loaded sheet-metal parts for welding on 10-hour shifts for about ten months (BMW) or eleven (Figure), both company-claimed. The reported totals do not show that the robot built vehicles or saved money; they describe vehicles supported during the deployment.


Humanoid robot vs industrial robot, cobot or AMR: which should a factory choose?

Choose by task variability, cycle time, integration need and comparative cost. IFR says humanoids complement existing robots, and Gartner says conventional options usually win on cost and throughput. A humanoid fits only where the requirement cannot otherwise be met, and the buyer should document why a conventional option fails before issuing a request for proposal.


What acceptance criteria should a humanoid robot pilot meet before scaling?

Define productive availability, successful cycles, first-pass quality, interventions, recovery time, integration cost and payback in advance, with pass, extend and terminate rules. All thresholds are buyer-set, because no published standard exists.


Should I buy humanoid robots in 2026–2027 or wait?

For most buyers, run a bounded paid pilot on a narrow task and defer production-scale capital. This is a proposed, conditional conclusion for manufacturing in 2026–2027, subject to the triggers in Section 8.6. A purchase case needs disclosed terms and measured results, and neither exists in the public record at the date of this report.

10. References & Strategic Sources

Every source is classified Verified or Company-claimed, following the research data. Verified means the document exists and says what is attributed, including filings, standards pages and statements made on the record. It does not mean the content is true. Company-claimed covers figures that originate with a manufacturer, OEM, vendor, union or analyst and are not independently audited. Sources follow a tier hierarchy: company filings and spec sheets first, then research bodies, analyst houses and trade press.

The original Step 2 label is kept in brackets or text beside each classification so a reader can see why a source was rated as it was. Tier 1 covers company filings, investor materials, releases and spec sheets. Tier 2 covers the Silicon Valley Robotics Center, IEEE Spectrum and IFR. Tier 3 covers Yole Group, Interact Analysis and Morgan Stanley or Goldman Sachs research. Tier 4 is trade press, used to confirm reported events and not as analysis. Items marked secondary or unverified were used only with attribution. Independent-analyst figures (McKinsey, Gartner, Goldman Sachs, Interact Analysis) carry the research data's company-claimed classification and appear in the text as analyst estimates or forecasts. Five sources that the research data lists without a Verified or Company-claimed label are shown as Company-claimed, following its rule that weaker labels map to company-claimed. The Journal's own earlier post and six sources the research data excluded, paywalled or did not retrieve are omitted. Every URL must be opened and confirmed before publication, and any source that cannot be opened should be removed. The method is described on the Research Methodology and Editorial Standards pages.



This report is backed by authoritative research, independent verification, and structured analytical methodology.


Decision for the buyer: before approving any capital, define the pilot scorecard (Section 8.3) and request the vendor evidence pack (Section 8.4). Proceed to a paid pilot on one narrow task only if the vendor accepts written metrics, data access, disclosed terms and exit rights, and defer production-scale capital until independently measured results exist.

bottom of page