The Hidden Integration Bill: What Humanoid Deployment Really Costs Beyond the Robot Price
Physical AI Journal Research Team — Independent analysis by Sekason Research Limited, United Kingdom
Physical AI Journal · physicalaijournal.org
Scope & Disclaimer
Scope statement: This report provides market and adoption intelligence for procurement decisions. It gives no legal, financial, investment, engineering or safety-certification advice, and no opinion on the safety, engineering soundness or technical conformity of any robot or deployment. It does not endorse any company or platform. Company-claimed figures are labelled and have not been independently verified by Physical AI Journal. Readers should not rely on this report in place of their own due diligence and professional advice.
Full terms are at physicalaijournal.org/disclaimer.
Figures from vendors, customers and analysts are reproduced as published and are not endorsed by Physical AI Journal or Sekason Research Limited. Nothing here is an offer, recommendation or solicitation to buy, sell or invest.
Five Signals Every Procurement Buyer Should Read Before Approving a Humanoid Budget
Run a pilot with written acceptance criteria before approving a humanoid budget, because humanoid robot deployment cost is visible at the robot-price and vendor-fee level but not at the level a buyer pays. Agility Robotics illustrates a ~$200,000 Digit purchase, ~$20,000 deployment fee and ~$36,000 a year in software and maintenance in SEC materials (company-claimed illustrative), yet none of the 4 named deployments discloses buyer-side integration cost or return on investment (ROI). Buy where utilisation and integration capacity can be evidenced, use Robots-as-a-Service (RaaS) where utilisation is uncertain, and wait where prices or evidence are not yet available.
1. The robot price is becoming measurable; the buyer's integration bill is not.
Agility's illustration gives a robot price, a vendor deployment fee and a recurring charge, more than any other vendor in this evidence set publishes (company-claimed illustrative). The same filing warns that unexpected integration complexity can raise implementation and support costs (VERIFIED company risk disclosure), yet no deployment discloses a buyer-side total.
2. A deployment fee and an integration cost are different numbers.
Agility's customer-facing fee is ~$20,000 for ownership and ~$25,000 for RaaS against a stated internal deployment cost of ~$15,000 (company-claimed / management model — not verified market pricing). BMW Group reports that integrating its Figure 02 robot involved production IT, occupational safety, process management and shop-floor logistics (customer-reported). The fee prices the vendor's work; the other functions are the buyer's. Assign an internal cost owner to each before approval.
3. Evidence stalls at announcement.
None of the 4 named deployments (BMW, Mercedes-Benz, GXO, Toyota Motor Manufacturing Canada) discloses buyer-side cost or ROI. BMW's reported activity, more than 30,000 X3 vehicles supported and ~1,250 operating hours, describes work done, not money spent (customer-reported). GXO's 100,000+ totes at ~98% accuracy is company-claimed and unaudited. Use reference customers to generate questions, not benchmarks.
4. Buy versus RaaS is visible only as a company model.
Agility illustrates ~$400,000 over 5 years for ownership and ~$500,000 for RaaS (company-claimed / management model — not verified market pricing). Both exclude buyer-side costs. Model the contract terms you would actually sign, not the headline totals.
5. The case for waiting has buyer-side support.
Mercedes-Benz board member Jörg Burzer has said cost will decide adoption. Reuters reported in August 2026 that industrial humanoids in China commonly cost around 300,000–500,000 yuan according to MERICS, and that Guotai Securities estimated ~160,000 yuan including maintenance would be needed for a 2-year payback (both analyst estimates). Define the price and evidence at which your approval would change.
Challenge first the layers that Agility's illustration leaves blank: IT and WMS integration, workflow engineering and supervision labour. Those are where the filing's own risk language points, and where BMW's account shows the work spreading across departments.
Signal | Evidence | Label | Budget implication |
Robot price measurable, integration bill not | Agility ~$200,000 / ~$20,000 / ~$36,000 a year; no buyer-side total disclosed | Company-claimed illustrative | Quote is line 1 of the budget |
Fee is not integration cost | Fee ~$20,000 / ~$25,000 vs ~$15,000 internal; BMW lists 4 functions | Company-claimed / management model — not verified market pricing; customer-reported | Name an internal owner per function |
Evidence stalls at announcement | 4 deployments, no cost or ROI disclosed | Customer-reported; company-claimed | References produce questions, not benchmarks |
Buy vs RaaS is a company model | ~$400,000 vs ~$500,000 over 5 years | Company-claimed / management model — not verified market pricing | Model signed terms |
The wait case is real | 300,000–500,000 yuan; ~160,000 yuan | Analyst estimate | Set a price and evidence trigger |
The Humanoid Market in 2026: Where Deployment Cost Evidence Actually Exists
Roughly 7,000 full-size humanoid robots were sold worldwide in 2025 (VERIFIED, from the International Federation of Robotics, IFR), but most applications remain specialised and many still depend on human teleoperation. Vendor order and production figures, including Agility's more than $300 million of Digit 5 orders and Figure AI's more than 350 Figure 03 robots, are company-claimed and describe orders or output, not productive fleets. Deployment cost evidence exists mainly in Agility's SEC materials and BMW's customer-reported activity data.
The IFR baseline, published on 30 September 2026 in its World Robotics 2026 service-robot release, is approximately 7,000 full-size humanoids sold in 2025 for commercial and professional use (VERIFIED, Tier 1). IFR also reports that most applications remain specialised, that many robots still require human teleoperation, and that high training and maintenance costs and weak industrial business cases continue to hold adoption back. For a buyer, the baseline says the installed base is small and the cost evidence thin. Teleoperation dependence also changes the budget shape: supervision labour is a continuing operating cost, not a one-off launch cost.
Agility names manufacturing, logistics and distribution as its first target markets, citing large labour pools, significant labour spending, repetitive physical work and environments built around people (company-claimed). It says Digit is deployed with GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada (company-claimed). Only GXO's pilot is confirmed by a customer disclosure in this evidence set; the other references rest on Agility's own materials. A target market is a vendor's sales plan, so it tells a buyer where the vendor expects to sell, not where the economics have been shown to work.
Integration is now sold as a product. Agility's stack combines hardware, Agility Arc deployment and fleet software, deployment services, maintenance, customer enablement and integration with WMS, WES, MES and existing automation (company-claimed). Reuters reported that Humanoid would assist Schaeffler with integration into existing factory lines (company-claimed). A line item sold by a vendor can be itemised, challenged and negotiated by a buyer.
Scale announcements describe intent, not delivery. Agility reported more than $300 million of multi-year Digit 5 orders covering 1,000 robots under a 3-year RaaS structure, subject to contractual milestones (company-claimed, June 2026). The figure describes orders subject to contractual milestones, not delivered or productive robots. Figure reports more than 350 Figure 03 robots produced and output raised from 1 robot a day to 1 robot an hour, a claimed 24× increase, with initial capacity of 12,000 robots a year and a target of 100,000 over 4 years (company-claimed). Humanoid, a robot maker, plans 1,000–2,000 robots at Schaeffler through 2032, with initial rollout from December 2026 to June 2027 and no contract value disclosed (company-claimed, Reuters, May 2026).
Agility unveiled Digit 5 on 15 September 2026 after reporting more than 65,000 hours of operational experience across earlier Digit deployments (company-claimed). That figure measures accumulated vendor experience. It does not measure cost, productivity or the buyer's integration effort.
Use one test on every number in this market: does it describe money a buyer pays, work a robot did, or something a vendor intends? Only the first answers a budget question, and it is the rarest.
Three numbers must stay separate: units sold (VERIFIED), contracted orders (company-claimed) and productive deployed fleets, which no vendor in this evidence set discloses. A buyer reading a vendor presentation should ask which of the three each slide describes. Vendor scale figures show capacity ambitions and confidence. They do not show that the integration bill has been solved.
Item | Figure | Label | Tier | Evidence stage |
Full-size humanoid sales, 2025 | ~7,000 units | VERIFIED | Tier 1 | Market baseline |
Digit 5 multi-year orders | >$300M; 1,000 robots; 3-year RaaS | Company-claimed | Tier 1 | Announcement (milestone-based) |
Figure 03 output | >350 robots; 1 robot/hour | Company-claimed | Tier 3 | Production claim (manufacturing, not deployment) |
Figure BotQ capacity and target | 12,000 robots/year; 100,000 over 4 years | Company-claimed | Tier 3 | Announcement (target) |
Humanoid at Schaeffler | 1,000–2,000 robots through 2032 | Company-claimed | Tier 2 | Announcement |
Digit operating experience | >65,000 hours | Company-claimed | Tier 3 | Vendor-reported |
What Happened After the Press Release: Deployment Evidence by Stage
Reported operating activity exists; reported economics do not. BMW reports (customer-reported) that Figure 02 supported production of more than 30,000 X3 vehicles, moved more than 90,000 components and logged ~1,250 operating hours over 10 months. No named deployment publicly discloses buyer-side cost, internal engineering hours or ROI, so none can serve as a cost reference for a budget, and every claim should be dated and staged.
The PAJ Evidence Pipeline places every deployment on 4 stages: Announcement, Pilot, Commercial Deployment and Production. Placement below rests on the Journal's research data; where it is the Journal's judgement, the table says so.
BMW's Figure 02 deployment at its Spartanburg plant in South Carolina is a Pilot. The 2025 pilot reported results in February 2026: more than 30,000 X3 vehicles supported over 10 months of 10-hour shifts, Monday to Friday, with more than 90,000 components moved and ~1.2 million steps completed over ~1,250 operating hours (customer-reported). In June 2026 Figure 03 arrived at Spartanburg, which Figure says moved beyond sheet-metal pick-and-place into sequencing and dynamic material manipulation (company-claimed). Operating hours are hours worked, not hours productive net of intervention and downtime, and the second measure is the one a return model needs.
Mercedes-Benz is also at Pilot: Reuters reported a handful of Apptronik Apollo units under test. GXO announced a pilot in December 2023, and its metrics come from Agility's September 2026 investor materials (company-claimed). Toyota Motor Manufacturing Canada signed a RaaS agreement in February 2026 after a pilot described as successful (company-claimed), which makes it a Commercial Deployment of undisclosed scale. Schaeffler's planned rollout is an Announcement. No deployment in the evidence set can be placed at Production on independently documented evidence.
What none discloses is the same in every case: integration cost, internal engineering hours, training cost, site modification, annual support and intervention labour. A complete buyer-side cost and an audited financial outcome do not appear anywhere in the evidence set. A reference deployment becomes decision-grade only when it discloses, at a minimum:
the buyer's total spend from signature to steady operation, split by layer;
productive hours after intervention and downtime;
supervision and teleoperation labour;
recurring charges actually paid;
status at 6 and 12 months;
whether the fleet expanded.
Movement between stages is the evidence worth tracking. An Announcement becomes a Pilot only when a named customer reports operating activity; a Pilot becomes a Commercial Deployment only when a commercial agreement and recurring operation follow; and Production requires sustained, disclosed operation at scale. The Toyota agreement shows the second step claimed but unquantified, and the GXO figures show operation claimed by the vendor but not audited. The Schaeffler plan shows the first step still pending, with initial rollout scheduled from December 2026 to June 2027 (company-claimed). For a procurement team, the practical rule is to date every claim: a robot count without a date and a stage is an intention, however large. A buyer's own approval memo should carry the same discipline, naming the stage each supporting reference has actually reached, the date and the source that says so.
Agility's disclosed Customer Acceleration path gives a stage model for the buyer's own timeline: proof of technology, proof of concept, RaaS pilot, then commercial deployment, an illustrative path of 6 months or more to scaled deployment (company-claimed). Robot delivery is therefore not operational deployment. Ask each vendor for reference customers at Commercial Deployment stage, and treat anything that cannot show the items above as a pilot.
Deployment | Stage | Disclosed metrics | Not disclosed | Label |
BMW × Figure 02, Spartanburg | Pilot (Journal placement) | >30,000 X3 vehicles; >90,000 components; ~1.2M steps; ~1,250 hours; 10 months | Cost, ROI, internal engineering hours | Customer-reported |
BMW × Figure 03, Spartanburg | Pilot (Journal placement) | Arrival June 2026; sequencing and dynamic material manipulation | All operating and cost metrics | Deployment verified; performance company-claimed |
Mercedes-Benz × Apollo | Pilot | Handful of units; component transport and quality checks | Productivity, ROI, cost | Verified deployment; outcome undisclosed |
GXO × Digit, Flowery Branch | Pilot (Journal placement) | 100,000+ totes; ~98% accuracy | Cost, integration spend, audit | Company-claimed |
Toyota Motor Manufacturing Canada × Digit | Commercial Deployment (Journal placement) | RaaS agreement, February 2026 | Scale, price, outcome | Company-claimed |
Humanoid × Schaeffler | Announcement | 1,000–2,000 robots through 2032 | Contract value | Company-claimed |
Humanoid Robot Deployment Cost: Building the Integration Bill Layer by Layer
Beyond the purchase price, the disclosed vendor items are a deployment fee and recurring software and maintenance: Agility's illustration puts them at ~$20,000 and ~$36,000 a year on a ~$200,000 robot (company-claimed illustrative). Site readiness, IT integration, workflow engineering, training and supervision costs fall on the buyer and are not publicly disclosed by any named deployment. They must be priced from the buyer's own site, with a named internal owner for each layer.
4.1 The Cost Stack
The path from signed purchase order to productive operation breaks into 11 layers.
Two classifications decide the budget: whether a layer is vendor-charged or buyer-incurred, and whether the evidence set discloses a figure for it. Only 4 of the 11 layers carry any public figure, and for software and maintenance the figure is a single combined number. The other 7 depend on the buyer's site, so no vendor illustration or third-party calculator can substitute for the buyer's own estimate. No dollar value is assigned to any undisclosed layer in this report. Layers 3 to 7, 10 and 11 are the ones where scope moves with the site. Under RaaS, layer 1 becomes a recurring charge, yet Agility's illustration keeps a separate deployment fee. Assign each undisclosed layer to a named internal owner in IT, operations, process management or logistics before capital is approved, and record which layers the vendor's quote already covers.
Layer | Who pays | Timing | Disclosed in evidence set? |
1. Robot acquisition | Vendor-charged | One-time (monthly under RaaS) | Yes: Agility ~$200,000 ownership (company-claimed illustrative); Figure and Apptronik prices not located |
2. Vendor deployment fee | Vendor-charged | One-time | Yes: ~$20,000 ownership, ~$25,000 RaaS (company-claimed / management model — not verified market pricing) |
3. Site readiness | Buyer-incurred | One-time | No cost; activity named in Agility's risk disclosure |
4. IT, WMS, WES and MES integration | Buyer-incurred; split with vendor undisclosed | One-time | No cost; BMW names production IT involvement |
5. Workflow engineering | Buyer-incurred | One-time | No |
6. Training and teleoperation | Buyer-incurred | One-time and recurring | No cost; IFR names training cost as a barrier |
7. Safety and infrastructure adaptation (cost layer only) | Buyer-incurred | One-time | No cost; BMW names occupational safety involvement |
8. Fleet software | Vendor-charged | Recurring | Combined with maintenance: ~$36,000 a year (company-claimed illustrative) |
9. Maintenance | Vendor-charged | Recurring | Combined with software: ~$36,000 a year (company-claimed illustrative) |
10. Downtime and ramp-up | Buyer-incurred | One-time and recurring | No; 6+ month illustrative path |
11. Ongoing optimisation | Buyer-incurred | Recurring | No |
4.2 What Agility's Filing Discloses
Agility's economics appear in an investor presentation and SEC filing connected with its proposed combination with Churchill Capital Corp XI (September 2026), so the audience is investors. Under ownership, a ~$200,000 robot, a ~$20,000 deployment fee and ~$36,000 a year in software and maintenance give roughly $400,000 over 5 years (company-claimed illustrative). Each of these is a company-claimed illustrative management assumption, not verified customer contract pricing. The filing itself says actual contracts are subject to negotiation (company-claimed illustrative). Under RaaS, ~$8,500 a month plus a ~$25,000 deployment fee is stated by the company as roughly $500,000 over 5 years (company-claimed / management model — not verified market pricing).
The ownership total reconciles: 5 × $36,000 + $200,000 + $20,000 = $400,000 (Editorial Estimate). The RaaS total is stated only as approximately $500,000 (company-claimed / management model — not verified market pricing), so a buyer should model contract terms, not headline figures.
The contrast between the two models is a choice about timing. Ownership commits ~$220,000 before the first productive shift; RaaS front-loads only the ~$25,000 deployment fee, a difference of $195,000 in upfront cash that leaves out the first monthly payment (Editorial Estimate). The stated 5-year totals put RaaS ~$100,000 higher (Editorial Estimate), so deferring capital has an illustrated price that a buyer should compare with its own cost of capital.
Two assumptions in the illustration deserve testing before it is used as a template: the 5-year horizon, and recurring charges that stay constant across it. The evidence set states no escalation terms, so a buyer should ask for the escalation clause rather than assume there is none. Neither total includes any cost the buyer carries itself, such as internal engineering time, site changes or supervision, and neither says what happens to recurring charges if deployment is delayed.
Item | Ownership | RaaS | Label |
Robot | ~$200,000 | ~$8,500 a month | Company-claimed / management model — not verified market pricing |
Deployment fee | ~$20,000 | ~$25,000 | Company-claimed / management model — not verified market pricing |
Software and maintenance | ~$36,000 a year | Not itemised in evidence set | Company-claimed illustrative |
5-year total as stated | ~$400,000 | ~$500,000 | Company-claimed / management model — not verified market pricing |
4.3 Fee Versus Bill
Agility's own filing separates what the customer is charged from what deployment costs the vendor: an internal deployment cost of ~$15,000 against customer-facing fees of ~$20,000 for ownership and ~$25,000 for RaaS (company-claimed / management model — not verified market pricing). The spread is $5,000 and $10,000 respectively (Editorial Estimate). The fee prices the vendor's deployment work. It does not price the buyer's.
The buyer's side is where scope moves. Agility's risk disclosure states that each customer has different physical layouts, workflows, technology stacks, WMS, fleet orchestration, IT infrastructure and information-security requirements, and that integration delays and scope changes can increase implementation costs and support requirements (VERIFIED company risk disclosure). BMW's account points the same way: 4 functions, from production IT to shop-floor logistics, were involved early in its integration (customer-reported).
The consequence for a budget is direct. The fee is a vendor-side number, and the integration bill is the sum of every layer the fee does not cover. Three questions separate the two: what work the fee buys, who performs the remaining work, and who pays when the remaining work takes longer than planned. Before approving a budget, ask the vendor to state in writing what the deployment fee covers and what it excludes, then convert every exclusion into a named internal budget line. A fee that looks small beside a ~$200,000 robot (company-claimed illustrative) can still leave the largest cost with the buyer, and a quote that does not itemise it cannot be compared with another vendor's.
Item | Figure | Label | Your figure |
Customer deployment fee, ownership | ~$20,000 | Company-claimed illustrative | |
Customer deployment fee, RaaS | ~$25,000 | Company-claimed / management model — not verified market pricing | |
Agility internal deployment cost | ~$15,000 | Company-claimed illustrative | |
Spread, ownership / RaaS | $5,000 / $10,000 | Editorial Estimate | |
Site, IT, workflow, training, supervision | Not publicly disclosed | None in evidence set |
4.4 Sensitivity and Red-Team
Stress-testing the case starts with the vendor's fee: using only Agility's disclosed figures, a 25%, 50% or 100% overrun on the ~$20,000 deployment fee adds $5,000, $10,000 or $20,000, which is 1.25%, 2.5% or 5% of the ~$400,000 ownership illustration (Editorial Estimate). An overrun on the vendor's fee is not where the exposure sits.
Under the ownership illustration, ~$220,000 is committed before the first productive shift and ~$36,000 a year recurs whether or not the robot is used (Editorial Estimate from company-claimed figures). Utilisation therefore decides cost per productive hour far more than the deployment fee does. The exposure sits in the 7 undisclosed layers and in realised utilisation, which only the buyer can price from its own site data. The 8 stress tests in the table each name the budget line they hit and the decision they could flip, with a blank column for the buyer's own figure. If a single stress test moves the case from approval to rejection, that assumption is the one to resolve before capital is committed, through a pilot or a contract term. Run each test alone against the buyer's own base case first, then in pairs, because delayed deployment and integration overrun tend to arrive together. Record the first test that flips the decision and the evidence that would settle it.
Stress test | Direction of impact | Budget line affected | Decision it could flip | Your figure |
Lower utilisation | Fewer productive hours per unit of fixed cost | Acquisition, software, maintenance | Buy or RaaS to Wait | |
Integration overrun | Raises one-time buyer-incurred cost | IT, workflow, site layers | Buy to Pilot | |
Higher intervention hours | Raises supervision labour | Training and teleoperation | Buy to Wait | |
Delayed deployment | Defers benefit | Ramp-up | Buy to Pilot | |
Software price escalation | Raises recurring cost | Fleet software | Buy to RaaS or Wait | |
Downtime | Interrupts production | Downtime reserve | Pilot to Wait | |
Added engineering headcount | Raises internal cost | Workflow, IT layers | Buy to Pilot | |
Failed pilot | Strands pilot spend | Pilot and integration spend | Pilot to Wait |
Overrun on ~$20,000 fee | Added cost | Share of ~$400,000 | Label | Your figure |
25% | $5,000 | 1.25% | Editorial Estimate | |
50% | $10,000 | 2.5% | Editorial Estimate | |
100% | $20,000 | 5% | Editorial Estimate |
4.5 Buy, RaaS, Pilot or Wait
The 4 options win under different tests that the buyer can run. Buy fits when utilisation can be evidenced from the buyer's own site and internal engineering capacity exists for the undisclosed layers; ownership is lower over 5 years only on Agility's own assumptions (company-claimed). RaaS fits when utilisation is uncertain or capital is constrained, with the contract fixing the deployment fee and exit terms; Toyota Motor Manufacturing Canada took this route after its pilot (company-claimed). Pilot fits when integration scope or acceptance criteria cannot yet be priced, and Agility's own path begins with proofs and a RaaS pilot before commercial deployment (company-claimed). Wait fits when the case depends on prices or evidence not yet available. The tests are the same for every option: can utilisation be evidenced, can the 7 undisclosed layers be staffed, and can the contract move the downside to the vendor? Related Journal analysis covers whether to deploy in 2026 or wait until 2028.
The wait case has buyer-side support. Jörg Burzer, Member of the Board of Management, Production, Quality & Supply Chain Management, Mercedes-Benz Group AG, said on 18 March 2025:
"The cost will be decisive... when costs reach a two-digit thousand-dollar sum - which is absolutely possible - it will become very interesting."
Reuters reported in August 2026 that industrial humanoids in China commonly cost 300,000–500,000 yuan according to MERICS, while Guotai Securities estimated ~160,000 yuan including maintenance would give a 2-year payback against an assumed worker cost (analyst estimates, not observed market prices); see also the Journal's view of Chinese robot pricing. Set the price and evidence at which your answer changes, and write both into the approval memo so that the decision is revisited on facts rather than on vendor announcements.
Four Named Deployments: What Was Disclosed, What Was Not
BMW provides the only quantified customer-side data set: more than 30,000 X3 vehicles supported, more than 90,000 components moved and ~1,250 operating hours over 10 months (customer-reported). Mercedes-Benz reports testing, GXO's results come from Agility (company-claimed) and Toyota Motor Manufacturing Canada reports a successful pilot followed by a RaaS agreement (company-claimed). None of the 4 discloses buyer-side cost or ROI. That absence is the finding.
BMW Group × Figure 02, Spartanburg, USA
Figure 02 worked at BMW's Spartanburg plant in South Carolina, a Pilot on the Journal's placement, with results reported in February 2026. BMW says the robot supported production of more than 30,000 X3 vehicles over 10 months, working 10-hour shifts Monday to Friday, and moved more than 90,000 components over ~1,250 operating hours and ~1.2 million steps (customer-reported). The integration evidence is the richest in the set: BMW says production IT infrastructure, occupational safety, production process management and shop-floor logistics were involved early, with standardised interfaces into its Smart Robotics ecosystem.
Michael Nikolaides, Senior Vice President, Production Network, Supply Chain Management, BMW Group, said on 27 February 2026:
"Our aim is to be a technology leader and to integrate new technologies into production at an early stage."
Cost, internal hours and ROI are undisclosed, and Figure 03's June 2026 arrival has no operating data yet. Ask any vendor which of those 4 functions its quote assumes the buyer will staff. For a buyer, the BMW account (customer-reported) is a template for the internal side of the bill: 4 named functions are 4 budget owners. Its numbers show activity at a named site over a stated period, which is more than most deployments disclose, yet they cannot be converted into cost per productive hour without BMW's cost and intervention data. Treat BMW as the standard for what a reference should disclose, not as evidence of return.
Mercedes-Benz × Apptronik Apollo, Berlin-Marienfelde and Kecskemét
Mercedes-Benz is testing Apollo at its Digital Factory Campus in Berlin-Marienfelde and in Kecskemét, Hungary, for component transport and quality checks, from 2025 onward. Reuters reported a handful of units, which places it at Pilot. Employees have used teleoperation and augmented reality (AR) to transfer production knowledge to the robot, with the aim of moving toward autonomous operation (customer-reported). That is a labour input the buyer pays for, and no cost is disclosed. No verified productivity or ROI outcome has been published; the confirmed fact is testing, not productive scale. The absence of an ROI statement is information about the stage, not a negative result.
Jörg Burzer's statement on cost, quoted in the economics section, sets the commercial bar from the buyer's side. Ask what supervision hours a pilot consumed and who supplied them, because that figure is the one most likely to be missing from a vendor's summary. The case is also a reminder that a pilot can be a long-running test rather than a step toward scale. The confirmed facts are 2 sites, a small number of units and a knowledge-transfer method that depends on people; how long testing has run and what would trigger a decision to scale are not disclosed.
GXO × Agility Digit, Flowery Branch, Georgia
GXO announced its pilot in December 2023 at the SPANX facility in Flowery Branch, Georgia, with Digit moving totes from autonomous mobile robots to conveyors, a repetitive material-handling task (customer-reported). Agility's September 2026 investor materials report 100,000+ totes at ~98% accuracy (company-claimed, not independently audited). The placement is Pilot, on the Journal's judgement. The metrics describe throughput and accuracy; integration spend, support cost, supervision labour and any effect on staffing are undisclosed. Two sources are involved, a customer pilot announcement and a vendor investor claim, and only the second carries the numbers. Ask for the same figures from GXO directly, not from the vendor, and treat the vendor's accuracy figure as a claim until GXO or an auditor confirms it. The case also separates a pilot announcement from a sustained operation: the December 2023 announcement is dated and customer-sourced, while the throughput and accuracy figures are recent and vendor-sourced. A buyer should request the record in between, including how many units ran and what support they required.
Toyota Motor Manufacturing Canada × Agility Digit, Ontario
Toyota Motor Manufacturing Canada signed a RaaS agreement in February 2026 following a pilot described as successful, covering manufacturing, supply-chain and logistics work (company-claimed). It is a Commercial Deployment of undisclosed scale. No quantitative productivity or ROI data was located, and neither the price under the agreement nor the number of robots is disclosed. A successful pilot is a statement, not a result. Ask what acceptance criteria defined success, who measured them and what the agreement says if a later phase misses them. The case shows RaaS used as the commercial step after a pilot, which is consistent with Agility's illustrated path (company-claimed). It does not show how the agreed monthly price compares with the ~$8,500 illustration (company-claimed / management model — not verified market pricing), and the evidence set contains no Toyota-side statement.
The 4 cases show a split. Integration-burden evidence is richest at BMW, on the customer side; cost evidence is richest at Agility, on the vendor side. No case supplies both, which is why a buyer must obtain both from its own vendor and its own site before approving capital. What a buyer can state with confidence from all 4 cases is limited to activity, announcements and agreements; cost, productivity and return remain to be established at the buyer's own site. Any vendor reference that cannot be traced to a customer statement should be classed as company-claimed.
Customer | Platform | Stage | Disclosed | Undisclosed | Label |
BMW Group | Figure 02 | Pilot (Journal placement) | Operating activity; functions involved | Cost, ROI | Customer-reported |
Mercedes-Benz | Apollo | Pilot | Handful of units; teleoperation and AR knowledge transfer | Productivity, ROI, cost | Verified deployment |
GXO | Digit | Pilot | 100,000+ totes; ~98% accuracy | Cost, audit | Company-claimed |
Toyota Motor Manufacturing Canada | Digit | Commercial Deployment | RaaS agreement | Scale, price, outcome | Company-claimed |
Where Humanoid Deployment Budgets Break: Friction, Risk and Unresolved Issues
The largest financial risks are integration scope that varies by site, ramp-up longer than planned, supervision and teleoperation labour, downtime, and contract terms that leave overruns with the buyer. Agility's own SEC risk disclosure names integration delays and scope changes as cost drivers (VERIFIED company risk disclosure), and IFR names training and maintenance costs as adoption barriers (VERIFIED). Each risk below is paired with the contract lever that can move it.
The 7 risks below are ordered by the Journal's judgement of budget exposure (Editorial analysis), with integration scope first because it touches the most cost layers.
Integration scope varies by site.
Agility's disclosure says layouts, workflows, WMS, fleet orchestration, IT infrastructure and information-security requirements differ for every customer, and that scope changes raise implementation cost and support requirements (VERIFIED company risk disclosure). BMW's 4 involved functions (customer-reported) show how many budgets one deployment touches. Because a vendor cannot know the buyer's site before surveying it, any quote issued ahead of a site survey carries scope risk that someone must hold. The lever is itemised scope with change-control terms that state who pays for each change.
Ramp-up takes time.
Agility's illustrative path runs 6 months or more from proof of technology to scaled deployment (company-claimed). Every month of ramp-up is a month in which one-time costs have been spent and productive output is still building. Establish when recurring charges start relative to acceptance, because benefit arrives after the charges may begin. The lever is payment tied to acceptance milestones rather than to delivery.
Humans stay in the loop.
IFR reports that many humanoid applications still require teleoperation, and Mercedes-Benz uses teleoperation and AR to teach its robot (customer-reported). Supervision labour is therefore a recurring operating cost that no public price list includes. The lever is a clause stating who supplies and pays for supervision, plus reporting of intervention hours so the buyer can see the labour the robot still consumes.
Downtime is a cost question.
IEEE Spectrum reports that industrial customers are highly sensitive to downtime and cites reliability expectations around 99.99% where an interruption can stop production (VERIFIED, independent analysis). This report offers no view on any platform's performance; the budget issue is who bears the cost of an interruption and how it is measured. The lever is uptime and support commitments measured at the process level, with remedies.
Manufacturing scale and supply constrain delivery.
Agility's SEC risk factors cite manufacturing difficulties, constraints on actuators, sensors, batteries and computing, skilled-personnel needs and quality control, and warn that deployment delays can increase costs (VERIFIED company risk disclosure). A vendor ramping production is also a vendor whose delivery dates can move. The lever is delivery schedules with remedies and spare-parts commitments. See the Journal's analysis of vendor risk in 2026.
Prices may move.
Reuters reported analyst estimates of 300,000–500,000 yuan for industrial humanoids in China (MERICS) and ~160,000 yuan including maintenance for a 2-year payback (Guotai Securities); neither is an observed market price. Buying early risks paying ahead of price movement, and a robot bought at today's price may later be compared with cheaper alternatives. The lever is exit, upgrade and obsolescence terms.
Contracts decide who pays.
As Editorial analysis, a buyer should establish who bears cost when integration runs long, uptime misses target, software changes break workflows, hardware becomes obsolete, spares are delayed, support is slow or pilot acceptance criteria are not met. The evidence set does not describe what vendors currently offer on any of these, so the answer must be obtained in negotiation. The Journal's piece on risk allocation in humanoid deployments covers the subject further.
Unresolved issues remain. No named deployment discloses intervention rates, utilisation or audited outcomes, and no vendor in the evidence set publishes a buyer-side cost total. Until they do, each risk above has to be priced from the buyer's own site rather than from a reference. After this section, a buyer should know which risks the budget prices, which a contract can move and which remain the buyer's own.
Risk | Evidence source | Label | Budget line affected | Contract lever |
Integration scope varies by site | Agility SEC risk disclosure; BMW | VERIFIED company risk disclosure; customer-reported | IT, workflow, site layers | Itemised scope; change control |
Ramp-up length | Agility Customer Acceleration path | Company-claimed illustrative | Ramp-up, deferred benefit | Charges tied to acceptance |
Supervision and teleoperation | IFR; Mercedes-Benz | VERIFIED; customer-reported | Training and teleoperation | Responsibility; intervention reporting |
Downtime | IEEE Spectrum | Tier 2 reporting | Downtime reserve | Uptime and support remedies |
Manufacturing and supply | Agility SEC risk factors | VERIFIED company risk disclosure | Spares, delivery timing | Delivery and spares terms |
Price movement | Reuters; MERICS; Guotai Securities | Analyst estimate | Acquisition, obsolescence | Exit and upgrade terms |
Contract cost allocation | Editorial analysis | Editorial Estimate | All layers | Itemised remedies |
Comparing Platforms by Commercial Disclosure and Deployment Burden Evidence
Public evidence does not support a defensible ranking of humanoid platforms by deployment burden. Only Agility discloses deployment-fee economics (company-claimed illustrative), and only BMW discloses which customer functions an integration involved (customer-reported). For Figure 03, Apptronik Apollo and Tesla Optimus no enterprise price was located, so a buyer cannot budget against them from public sources. Ranking on that basis would present missing data as a result.
The matrix compares what each vendor discloses commercially, not what any robot can do. Agility's Digit has a disclosed ownership and RaaS model and a deployment fee. Agility also says Digit integrates with autonomous mobile robots, WMS, WES and MES (company-claimed); that is a statement about the vendor's integration offer, not an assessment of how it performs at any site.
Figure and Apptronik are visible through their customers, BMW and Mercedes-Benz, rather than through price lists. A buyer must ask each for the commercial disclosure Agility has put in its filing: purchase price, commercial model, deployment fee, recurring charges and what the fee excludes. Unitree's G1 has a published online price of $13,500, with shipping and customs extra (VERIFIED published price), and 1X's NEO is a home product offered at $20,000 or $499 a month (company-claimed published offer). Both are pricing-model context only; they are not comparable with enterprise deployment economics and are excluded from the deployment-burden comparison. Tesla's Optimus has no comparable public enterprise price.
"Not located" is a finding. For every platform where the matrix shows it, a buyer should require the vendor to supply the missing items before shortlisting, since at least one vendor has shown such disclosure is possible. A ranking would need comparable cost and outcome data across platforms. The evidence set holds cost disclosure for 1 platform and customer-side integration evidence for 1 customer.
Company-published specifications, shown for orientation only and not compared, include a 23 kg / 50 lb payload for Digit, a 20 kg payload and 5-hour runtime for Figure 03, and a 55 lb lift and four-hour battery operation for Apollo (company-claimed). They are not evidence of fitness for any application.
Before shortlisting, ask every vendor for the same 6 items: purchase price and commercial model; deployment fee and exclusions; recurring charges and escalation terms; named customer deployments with permission to contact; the customer functions involved in integration; and any acceptance criteria used in pilots. Compare vendors on how many of the 6 they can answer in writing, and treat the count as a measure of disclosure, not of quality. Where a vendor declines to disclose, record that as a risk to the budget, because an undisclosed layer is an unpriced one.
Platform | Published price | Commercial models | Deployment fee | Named customer deployments | Integration-burden evidence | Evidence stage | Label |
Agility Digit 5 | ~$200,000 ownership illustration | Ownership; RaaS ~$8,500/month | ~$20,000 / ~$25,000 | GXO; Toyota Motor Manufacturing Canada | Agility SEC risk disclosure | Commercial Deployment (Journal placement) | Company-claimed / management model — not verified market pricing |
Figure 03 | Not located | Not located | Not located | BMW Spartanburg | BMW functions list (Figure 02) | Pilot | Deployment verified; claims company-claimed |
Apptronik Apollo | Not located | Not located | Not located | Mercedes-Benz (testing) | Teleoperation and AR knowledge transfer | Pilot | Verified deployment |
Unitree G1 | $13,500 online; shipping and customs extra | Online sale | Not located | Not located | Not located | Not assessed | VERIFIED published price |
Tesla Optimus | Not publicly priced | Not located | Not located | Not located | Not located | Not assessed | No verified price |
1X NEO | $20,000 or $499/month (home product) | Ownership; subscription | Not applicable | Not located | Not located | Excluded (home) | Company-claimed published offer |
What an Automation/Procurement Buyer Should Do Before Approving Capital
Run a pilot with written acceptance criteria unless utilisation and internal integration capacity can already be evidenced, and negotiate the contract terms below before committing capital. Buy when both are evidenced, use RaaS when utilisation is uncertain, and wait when the case depends on prices or evidence not yet available. This is a decision framework, not investment advice. Stage the approvals so capital follows evidence.
Decision rule. Start from the pilot. If the buyer can evidence utilisation and has capacity for the 7 undisclosed layers, ownership is the candidate; if utilisation is uncertain, RaaS with exit terms is; if neither can be shown, wait. The Journal's guidance on when not to buy a humanoid robot is a useful counterweight to a case built on vendor illustrations.
Stage the approval. Approve capital in the same 4 steps as Agility's illustrative path (company-claimed): a small amount for the site survey and proof, a larger amount for the pilot against written acceptance criteria, and fleet capital only against measured results from the buyer's own site. Define in advance what result at each gate stops the programme, because a gate with no stop condition is only a timetable.
Pre-approval checklist.
Name an internal owner and a budget line for each of the 11 cost layers.
Obtain the vendor's itemised list of exclusions and convert each into a budget line.
Agree site survey scope and acceptance criteria in writing before the pilot starts.
Set a utilisation target with a plan for how it will be evidenced.
Map the 4 stages to stage gates tied to payment.
Present a range to the approval committee. Show the vendor illustration as the floor, the buyer-side layers as named estimates with owners, and the stress tests with the first assumption that flips the decision. A committee that approves a single number is approving the vendor's assumptions. State in the memo which cost layers the vendor quote covers and which are assumed to sit with the buyer, so that no layer is unowned. Record the reasoning, not only the outcome.
Ask vendors for what the evidence set lacks: customer reference economics, intervention rates, and status at 6 and 12 months after deployment. A vendor that offers references should be asked to commit the same measures to the contract's reporting clause. Read the Journal's notes on the humanoid robot exit clause and vendor due diligence before negotiating. Revisit the decision when a reference deployment discloses buyer-side cost, a vendor publishes an enterprise price or the pilot misses its criteria. Assign the contract questions to named owners: procurement owns scope, price and exit terms, operations owns acceptance criteria and utilisation reporting, and IT owns integration scope and change control. The 10 contract questions below are procurement questions, not legal advice.
Term | Question to put to the vendor | Cost layer protected |
1. Deployment fee scope | What does the fee cover, and what does it exclude? | Vendor fee; site readiness |
2. Itemised exclusions | Which integration, training and site items are outside the quote? | All buyer-incurred layers |
3. Acceptance criteria and remedies | What measurable criteria end the pilot, and what follows if they are missed? | Pilot spend; ramp-up |
4. Uptime and support | What availability is committed, measured where, with what remedy? | Downtime |
5. Spare parts | Which spares are held, where, and with what lead time? | Maintenance; downtime |
6. Software change and price escalation | How are software changes tested and notified, and how are recurring fees capped? | Fleet software; workflow |
7. Teleoperation and supervision | Who supplies and pays for supervision and teleoperation hours? | Training and teleoperation |
8. Exit and obsolescence | What exit, buy-back or upgrade rights apply? | Acquisition; obsolescence |
9. Pilot-failure cost allocation | Who bears cost if pilot criteria are not met? | Pilot and integration spend |
10. Intervention and utilisation reporting | What data is reported, how often, and in what format? | Utilisation; optimisation |
What to approve now. Approve only a funded site survey and a pilot with written acceptance criteria, send the 10 contract questions in writing to every shortlisted vendor, and release fleet capital only when the buyer's own site data evidences utilisation and each of the 7 undisclosed layers has a named internal owner.
Executive FAQ: Humanoid Robot Deployment Cost and Procurement Decisions
No single figure answers the cost question for a humanoid deployment. The answers below are limited to disclosed evidence: vendor illustrations are company-claimed, customer figures describe activity rather than economics, and buyer-side integration costs are not publicly disclosed. Each answer states what is known, what is not, and the decision it supports for an Automation/Procurement Buyer. The cost question is answered in layers: the vendor illustration first, then the buyer's own undisclosed costs.
How much does it cost to deploy a humanoid robot beyond the purchase price?
Agility's illustration shows a ~$20,000 deployment fee for ownership (~$25,000 for RaaS) and ~$36,000 a year in software and maintenance (company-claimed / management model — not verified market pricing). Those are the only deployment-related prices in this evidence set. Site readiness, IT integration, workflow engineering, training and supervision costs borne by the buyer are not publicly disclosed by any named deployment, so they must be estimated from the buyer's own site and contract. Ask the vendor to list in writing what its fee excludes.
What integration costs should I budget for when deploying humanoid robots in a factory or warehouse?
Budget for site assessment, WMS, WES and MES integration, workflow engineering, training and teleoperation, safety and infrastructure adaptation, and ramp-up. Agility's SEC risk disclosure names site assessment, workflow integration, WMS integration, training and ramp-up as deployment steps (VERIFIED company risk disclosure). BMW reports that production IT, occupational safety, process management and shop-floor logistics were involved early (customer-reported). No dollar value for these layers is publicly disclosed, so each needs a named internal owner and a budget line before approval.
What is the total cost of ownership of a humanoid robot over five years?
Agility illustrates ~$400,000 over 5 years for ownership and ~$500,000 for RaaS (company-claimed / management model — not verified market pricing). Both exclude buyer-side integration, internal labour and downtime. Ownership is lower than RaaS over 5 years only on Agility's own assumptions, and the filing itself says actual contracts are subject to negotiation (company-claimed illustrative). A buyer's TCO is therefore the vendor illustration plus its own undisclosed layers, priced from its own site and contract terms.
Should I buy a humanoid robot, use RaaS, run a pilot or wait?
Run a pilot with written acceptance criteria unless utilisation and internal integration capacity can already be evidenced. Buy when both are evidenced, use RaaS when utilisation is uncertain, and wait when the case depends on lower prices; Mercedes-Benz board member Jörg Burzer has said cost will decide adoption, and Reuters reported analyst estimates of 300,000–500,000 yuan for industrial humanoids in China. This is a decision framework, not advice. Write the price and evidence that would change the decision into the approval memo.
What have BMW, GXO and Mercedes-Benz actually achieved with humanoid robots?
BMW reports that Figure 02 supported production of more than 30,000 X3 vehicles, moved more than 90,000 components and logged ~1,250 operating hours over 10 months (customer-reported). Agility reports that GXO's Digit moved 100,000+ totes at ~98% accuracy (company-claimed, not independently audited). Mercedes-Benz is testing a handful of Apollo units. None of the three discloses cost or ROI, so use these cases to frame questions for vendors, not as benchmarks.
What should I put in a humanoid robot contract before signing?
Put 5 terms first: itemised deployment scope with named exclusions, acceptance criteria with remedies, uptime and support commitments, limits on software changes and price escalation, and exit and obsolescence terms. These map to the cost layers a buyer carries if the vendor's illustration proves optimistic. Assign each term to a named internal owner before negotiation starts. They are procurement questions, not legal advice; the full list of 10 appears in the recommendations section.
References & Strategic Sources
Sources are graded in 3 tiers used in this report. Tier 1 covers filings, IFR data and customer or company investor disclosures; Tier 2 covers Reuters, IEEE Spectrum and established business press; Tier 3 covers manufacturer press releases and specification pages, used only where primary commercial information is unavailable. Every figure carries one of 3 labels: VERIFIED for independently verifiable data, COMPANY-CLAIMED for company or customer figures not independently validated, and Editorial Estimate for Journal arithmetic.
Evidence labelling key:
VERIFIED: independently verifiable from an official filing, customer disclosure or authoritative institutional source.
COMPANY-CLAIMED: a manufacturer or customer statement, or a company-generated performance or economic figure, not independently validated.
Editorial Estimate: arithmetic or judgement by Physical AI Journal from the figures cited, labelled wherever it appears.
Analyst estimate: a figure attributed to a named analyst or research body in independent reporting, not an observed market price.
To trace any figure, find its label, then its tier, then its source in the table below. A figure labelled COMPANY-CLAIMED from a Tier 3 source is a vendor statement, and one labelled VERIFIED from a Tier 1 source can be checked against the original document. Where sources conflict or are unavailable, the report says so rather than choosing silently. The tier and label of each source appear in the final column, and all dates are as published in the research data. The research cut-off is 5 October 2026. All 21 entries below come from the Journal's research data, and any entry without a supplied URL is marked "URL not located" instead of being given a placeholder link. The Journal's approach to sourcing is set out in its Research Methodology.
U.S. SEC / Agility Robotics: Agility/Churchill Investor Presentation & SEC filing (Sep 2026) – Classification: Company-claimed (illustrative economic model) · Tier 1 – https://www.sec.gov/Archives/edgar/data/2074973/000121390026071287/ea029548401ex99-2.htm
U.S. SEC / Agility Robotics: Risk Factors / Business Combination filings (Sep 2026) – Classification: Verified (company risk disclosure) · Tier 1 – URL not located
IFR: World Robotics 2026 Service Robots (30 Sep 2026) – Classification: Verified · Tier 1 – https://ifr.org/ifr-press-releases/news/global-robotics-race-korea-singapore-germany-japan-china
BMW Group: BMW Group to deploy humanoid robots in production in Germany for the first time (27 Feb 2026) – Classification: Verified (customer disclosure); performance figures company-claimed (customer-reported) · Tier 1 – https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time
BMW Group: Humanoid Robot in Leipzig (21 Sep 2026) – Classification: Verified (customer-reported) · Tier 1 – URL not located
Mercedes-Benz Group: AI and humanoid robots — Digital Factory Campus (18 Mar 2025) – Classification: Verified (customer disclosure) · Tier 1 – https://group.mercedes-benz.com/company/production/production-network/mbdfc-humanoid-robots.html
GXO Logistics: Industry-leading Digit pilot (6 Dec 2023) – Classification: Verified (customer disclosure) · Tier 1 – https://gxo.com/news_article/gxo-conducting-industry-leading-pilot-of-human-centric-robot/
Agility Robotics: Toyota Motor Manufacturing Canada commercial agreement (19 Feb 2026) – Classification: Company-claimed · Tier 3 – https://www.agilityrobotics.com/content/agility-robotics-announces-commercial-agreement-with-toyota-motor-manufacturing-canada
Agility Robotics: Digit 5 launch (15 Sep 2026) – Classification: Company-claimed · Tier 3 – URL not located
Figure AI: Ramping Figure 03 Production (29 Apr 2026) – Classification: Company-claimed · Tier 3 – URL not located
Figure AI: F.03 Arrives at BMW (30 Jun 2026) – Classification: Company-claimed · Tier 3 – URL not located
Apptronik: $935M Series A (11 Feb 2026) – Classification: Company-claimed · Tier 3 – URL not located
Reuters: Apptronik raises $520M (11 Feb 2026) – Classification: Verified (independent reporting) · Tier 2 – URL not located
Reuters: Mercedes-Benz tests Apollo (18 Mar 2025) – Classification: Verified (independent reporting) · Tier 2 – URL not located
Reuters: Humanoid to deploy up to 2,000 robots at Schaeffler (13 May 2026) – Classification: Verified (independent reporting) · Tier 2 – URL not located
Reuters: Chinese robots face commercial test (18 Aug 2026) – Classification: Verified (independent reporting; figures are analyst estimates) · Tier 2 – URL not located
IEEE Spectrum: Reality Is Ruining the Humanoid Robot Hype (2025) – Classification: Verified (independent analysis) · Tier 2 – URL not located
Unitree: Unitree G1 official store (Current 2026) – Classification: Verified (published price) · Tier 3 – URL not located
1X: NEO order/pricing page (Current 2026) – Classification: Company-claimed (published offer) · Tier 3 – URL not located
Apptronik: Apollo specifications (Current) – Classification: Company-claimed · Tier 3 – URL not located
Figure AI: Figure 03 specifications (Current) – Classification: Company-claimed · Tier 3 – URL not located
© 2026 Physical AI Journal, Sekason Research Limited. All rights reserved.
