Who Owns the Humanoid Deployment? How to Select and Contract a Systems Integrator Part I
The Industrial Buyer’s Framework for Integrator Selection, Contractual Accountability and Performance-Based Delivery
Physical AI Journal Research Team — Independent analysis by Sekason Research Limited, United Kingdom
Physical AI Journal (physicalaijournal.org) | Evidence as of 2 November 2026 | Primary reader: Automation/Procurement Buyer | Secondary reader: Operations Leader
THIS REPORT IS 2 PARTS
Scope & Disclaimer
This report provides market and adoption intelligence only. It gives no engineering, safety-certification, technical-conformity, legal, financial or investment advice or opinion, and Physical AI Journal does not evaluate or endorse any robot, platform or integrator. Acceptance indicators and scorecard weights are proposed buyer measures, not standards. Company-claimed figures are not independently verified. Readers should obtain independent professional advice before any procurement or contract decision.
Standards and regulation appear as dated status facts drawn from the sources listed in Section 11. Where a source is secondary legal commentary it is labelled as such and is not a substitute for advice from counsel. Nothing in this report is an offer, recommendation or endorsement.
Executive summary
This report supports one decision: whether to contract the humanoid maker, appoint an independent humanoid robot systems integrator, or run a multi-party arrangement, and how to tie payment to accepted results. None of the 4 humanoid deployments reviewed names a third-party integrator, so the buyer has to assign accountability itself. Multi-site commitments are already arriving: GXO and Agility announced an expansion to additional US sites on 26 October 2026 (company-claimed), and Humanoid (UK) and Schaeffler signed a phased agreement on 13 May 2026 with first deployments from December 2026 (announced target). The reviewed structures are vendor-direct robots-as-a-service (RaaS), where the manufacturer’s own software layer is described as carrying workflow integration (company-claimed); a captive-parent arrangement; or terms that are not disclosed. Ownership of the deployment outcome is being allocated privately, deal by deal, before standard accountability models exist.
For the Automation/Procurement Buyer, it shows how to qualify an integrator on evidence rather than references and when to defer appointment. It supplies a responsibility matrix, a weighted qualification scorecard, a commercial-model comparison and a contract-provision checklist. Every figure carries its evidence tier, including Agility’s illustrative RaaS price of about $8,500 per month per robot plus a $25,000 deployment fee (company-claimed), and the report states plainly where public evidence stops.
1. Executive Intelligence Synthesis
Direct answer. No reviewed case names an independent integrator. In the 4 deployments reviewed, the humanoid manufacturer holds the commercial relationship under vendor-direct RaaS, is the customer’s parent company, or has undisclosed terms, and no public source discloses how accountability for integration, acceptance and ongoing performance is allocated. The buyer therefore has to assign that ownership in its own contract. Every deployment figure reported to date is company-claimed; none has been independently verified.
Signal 1: Multi-site commitments are arriving before accountability models
GXO and Agility announced on 26 October 2026 an expansion to additional US sites, with robot numbers and terms undisclosed (company-claimed). Humanoid (UK) and Schaeffler signed a binding phased agreement on 13 May 2026, with deployment from December 2026 (announced target).
“Customers are moving beyond asking whether humanoid robots can work to deciding where they can scale them.”
(Peggy Johnson, CEO, Agility Robotics, SEC Form 425 communication, as reported by DC Velocity, 26 October 2026; a company-claimed executive statement, not evidence of outcomes.)
Decision implication: appoint the accountable party before signature, not after the pilot.
Signal 2: No reviewed deployment names a third-party integrator
GXO/Agility is vendor-direct RaaS, Hyundai/Boston Dynamics is captive-parent, BMW/Figure has undisclosed terms and Schaeffler/Humanoid (UK) is announced RaaS (all company-claimed). The only integrator-type entities found are Accenture and SAP, in a Vodafone Procure & Connect pilot with no robot maker named, and Robo Inc., launched on 10 August 2026 with no named deployment (company-claimed). This is a disclosure gap, not evidence that none exists.
Decision implication: have every bidder state in writing who holds each deliverable.
Signal 3: Outcome evidence is reported by the companies themselves
BMW reports that Figure 02 supported more than 30,000 BMW X3 vehicles over 10 months (company-claimed, customer-side), and GXO and Agility report more than 100,000 totes moved (company-claimed). Neither is audited. Interact Analysis states real-world applications were about 10% of 2025 production (21 May 2026).
Decision implication: release no payment on reported volumes; tie it to accepted tests (Section 3).
Signal 4: Regulation is moving while humanoid-specific standards are not final
The EU Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 (secondary legal commentary). ISO/CD 25785-1, for dynamically stable industrial mobile robots, is at committee-draft stage (ISO, accessed 2 November 2026), and no final humanoid-specific ISO, IEC, IEEE or A3 standard was found. The FCC barred new imports of foreign-made humanoid and quadruped robots on 28–29 July 2026 (PBS/AP), adding a country-of-origin screen. Dated status facts only.
Decision implication: name in the contract who holds conformity documentation; counsel decides its content.
Evidence labels used throughout (V1).
Label | Meaning |
No label, source named | Independent: IFR, McKinsey, NIST, ISO, regulators, existence of an SEC filing |
(company-claimed) | Company or executive figure, price or claim |
(company-claimed, customer-side) | A customer’s own outcome report |
(company-claimed; via secondary press) | Company figure read through trade press |
(announced target) | Company plan, capacity or forecast |
(secondary legal commentary) | Law-firm, insurer or vendor-blog commentary |
(low-reliability source) | Item flagged low reliability |
2. Platform / Market Landscape
Direct answer. The reviewed evidence shows 4 ownership and integration structures: manufacturer-retained robots-as-a-service, customer-owned robots with a manufacturer software and maintenance subscription, captive-parent deployment, and direct supply from the maker. Third-party integrators appear only in announcements without outcome data: the Accenture, SAP and Vodafone Procure & Connect pilot, whose robot maker is not named, and Robo Inc., which has no named deployment. All structure descriptions are company-claimed.
IFR counts almost 6,900 humanoids sold globally in 2025, with the overwhelming majority from Asia-headquartered companies and 171 humanoid manufacturers known to IFR (World Robotics 2026 Service Robots Executive Summary). IFR’s 30 September 2026 release gives about 7,000 full-size humanoids above 140 cm. The wording differs and the 2 figures are not reconciled here; a separate figure of around 13,000 for 2025 appears in The Next Web, attributed to an industry count (secondary). IFR also calls RaaS the dominant business model in the US, but that statement covers service robots, not humanoids specifically.
Ownership and integration structures as reported (V3). “None named” means no third-party integrator appears in the sources reviewed.
Manufacturer / entity | Structure as reported | Integrator named | Label |
Agility Robotics (Digit) | RaaS (Agility retains ownership; subscription includes Arc software and maintenance) or Ownership (customer buys; deployment fee, subscription, maintenance). Most v5 orders structured as RaaS | None named | (company-claimed) |
Boston Dynamics (Atlas) | Captive parent: Hyundai Motor Group is majority shareholder; deployments in Hyundai and Kia plants | None named | (company-claimed) |
Humanoid (UK) (HMND 01) | RaaS including fleet software, maintenance and 24/7 support; Bosch to act as contract manufacturer for Europe. Who operates the robots after deployment is not stated | None named | (company-claimed) |
Figure AI | BMW runs the deployment in its plant and Figure supplies and develops the robots, per a trade-press summary; commercial terms not disclosed | None found | (company-claimed; via secondary press) |
UBTech (Walker S2) | Direct supply described by UBTech as a “standardized, replicable turnkey solution” on its BrainNet platform; pricing and terms undisclosed | None named | (company-claimed) |
Apptronik (Apollo) | Pricing and contract terms not stated; Jabil’s role not found | None found | (company-claimed) |
GXO (host and operator) | Operational incubator programme; the Reflex Robotics deal is GXO’s second humanoid RaaS agreement after Agility; terms not found | None named | (company-claimed) |
Accenture, SAP, Vodafone Procure & Connect | Pilot announced 22 April 2026 at a Duisburg warehouse, scope autonomous visual inspection; Accenture designed and deployed the robot intelligence and operating framework, SAP led integration with SAP Extended Warehouse Management; robot maker not named | Integrator-type, software and workflow | (company-claimed) |
Robo Inc. (RoboStore) | US manufacturing and systems-integration company launched 10 August 2026; hardware-agnostic; no named deployment | Integrator | (company-claimed) |
Tesla (Optimus) | Internal use; no integrator or third-party customer; customer shipments planned for 2027 (announced target) | None | (company-claimed) |
Where an integration route is described, it runs through the manufacturer’s own software: Arc for Agility, Orbit for Boston Dynamics and fleet software for Humanoid (UK) (all company-claimed). Structure also decides who can be held to account: under RaaS the manufacturer keeps the asset, under Ownership the buyer holds it, and under a captive parent the group does. Investor-customers such as Schaeffler and Foxconn, both tied to Agility, combine buyer and capital-provider roles (Physical AI Journal analyst reading, not a sourced finding), so a buyer in that position should ask how exclusivity and pricing interact with the investment.
Decision implication: choose the structure first, because each one places integration risk with a different party.
3. Deployment Evidence: Reported Outcomes, Not Announcements
Direct answer. Acceptance should rest on sustained, baseline-referenced results, not on demonstrations or on cumulative volumes reported by a vendor. This report uses a 4-rung ladder: demonstration, paid pilot, accepted installation and sustained production. A demonstration shows a task can be done; an accepted deployment shows it was done to an agreed standard, measured against a baseline, over an agreed observation period. The indicators below are Physical AI Journal proposed buyer measures, and the buyer sets the thresholds.
The evidence ladder (V4). The case placement is a Physical AI Journal reading of reported status, not an assessment of performance.
Rung | What it shows | Reviewed case at this rung |
1. Demonstration | Task performed in controlled or training conditions | Hyundai/Boston Dynamics: training-centre status and announcement |
2. Paid pilot | Task performed on live work under a limited commercial or pilot agreement | BMW/Figure 02, described by BMW as a pilot (payment terms not disclosed); GXO/Agility at Flowery Branch, with an expansion announced |
3. Accepted installation | Results measured against agreed tests and accepted in writing | None of the reviewed cases discloses acceptance |
4. Sustained production | Agreed results held over an observation period at scale | None disclosed |
Schaeffler/Humanoid (UK) sits below rung 1: an agreement is signed and deployment begins from December 2026 (announced target). A signed agreement does not prove a successful deployment. Randy Levine, author, WhichBot.net, made the same point as commentary in RoboticsTomorrow on 23 October 2026: “Do not treat an announced pilot as proof of a scaled commercial rollout.”
Independent test data does not yet exist. The NIST humanoid baseline performance benchmark, published as a proposal on 20 April 2026 and last updated on 15 May 2026, has its test apparatus still in development, with no commercial robots tested and no results reported.
Tracker-level detail sits in the Humanoid Deployment Evidence Report 2026.
Claim treatment (V5). How each reported metric should be read.
Metric | Reported figure | Label | How to read it |
GXO totes moved at Flowery Branch | More than 100,000 (DC Velocity); other company materials give about 60,000 and over 125,000 | (company-claimed) | Volume at one site; no hours, uptime, error rate or cost per tote published; variants not reconciled |
BMW Figure 02 pilot | Over 10 months: more than 30,000 BMW X3, more than 90,000 components, about 1.2 million steps, roughly 1,250 operating hours | (company-claimed, customer-side) | BMW’s own report; Figure and trade press give 11 months, not reconciled |
Figure accuracy and cycle time | Placement accuracy above 99% per shift; 84-second cycle time, reported as a target by IIoT World | (company-claimed; via secondary press) | Vendor-sourced; target and reported measure are not distinguished across outlets |
Agility cumulative hours | 65,000 operating hours and 9 committed customer facility deployments (May 2026) | (company-claimed) | Cumulative across sites; never a single-site measure |
Schaeffler/Humanoid (UK) scale | 4-digit wheeled units by 2032 | (announced target) | Not a confirmed order |
The indicators work only if the contract fixes the baseline first, because without a pre-deployment baseline throughput and completion rates cannot be read. Autonomous completion matters because teleoperation share was not found in any reviewed case.
6 acceptance indicators (V6), each a Physical AI Journal proposed buyer measure; the buyer sets thresholds and the observation window
Indicator | Definition | Evidence source |
Task completion rate | Share of assigned tasks completed | Fleet logs reconciled to operator records |
Autonomous completion vs human intervention | Share of tasks completed without remote or on-site human help | Intervention log |
Availability in scheduled production hours | Time the system is available when production is scheduled | Fleet logs against the production schedule |
Recovery from failed tasks | Time and effort to return to task after a failure | Incident log |
Throughput against a defined baseline | Output compared with the pre-deployment baseline | Operator throughput records |
Sustained performance over the observation period | Results held across the agreed period, not a single best week | Weekly series over the window |
Decision implication: a demonstration must not trigger final payment; require rung 3 and rung 4 evidence, measured by the indicators above, before release.
4. Economics & ROI Analysis
Direct answer. The 4 commercial models allocate integration and performance risk differently: fixed fee leaves performance risk with the buyer, milestone-based payment ties release to accepted stages, performance-linked payment ties it to sustained results, and integrated RaaS bundles deployment and operating services into one subscription. Only the last has a publicly disclosed example: Agility’s RaaS at about $8,500 per month plus a $25,000 deployment fee (company-claimed). No independent integration fee benchmark was found.
The 4 commercial models
Only integrated RaaS has a publicly disclosed example; fixed-fee, milestone-based and performance-linked structures appear in no reviewed source.
Commercial-model matrix (V7). Physical AI Journal analysis of payment structures; the only disclosed example is Agility’s, and it is company-claimed.
Model | Who bears performance risk | What triggers payment | Incentive effect | Disclosed example |
Fixed fee | Buyer | Delivery of the agreed scope | Rewards finishing scope, not sustaining results | None |
Milestone-based | Shared, depending on how milestones are defined | Written acceptance of each stage | Rewards passing tests; risk that the tests measure a demonstration | None |
Performance-linked | Integrator or manufacturer | Measured results over an observation period | Aligns payment to outcome; needs an agreed baseline and metering | None |
Integrated RaaS with deployment and operating services | Depends on service terms; the provider retains robot ownership | One-time deployment fee plus recurring subscription | Subscription priced against a labour rate | Agility (company-claimed) |
Each model fails in a characteristic way. A fixed fee can be earned on installation day while results are still unconfirmed. Milestones protect the buyer only if the acceptance test measures production work rather than a demonstration. Performance-linked payment needs a baseline and metering that both sides trust. Bundled RaaS keeps the robot on the provider’s books, but the sources reviewed do not say who carries liability or uptime commitments. For any model, the buyer should ask what triggers each payment and what the buyer can withhold.
Where integration sits in disclosed pricing
Agility describes its Arc software as integrating Digit into facilities and workflows, coordinating robots and operators, and monitoring fleets across sites (company-claimed). Integration therefore sits inside the manufacturer’s subscription and deployment fee, and no separate integration line is disclosed. Agility also cautions that its unit economics rest on assumptions that could change (company-claimed; via secondary press).
Agility disclosed economics with arithmetic check (V8). S-4 figures were read through The Robot Report, 7 September 2026; deck figures come from the Form 425 investor presentation, May 2026.
Item | RaaS | Purchase (Ownership) | Label |
Upfront robot price | None; Agility retains ownership | About $200,000 | (company-claimed; via secondary press) |
Deployment fee | $25,000 | $20,000 | (company-claimed; via secondary press) |
Recurring fee | About $8,500 per month (about $102,000 per year), including Arc software and maintenance | Roughly $36,000 per year for software and maintenance | (company-claimed; via secondary press) |
Five-year revenue per robot, S-4 as reported | About $535,000 | About $400,000 | (company-claimed; via secondary press) |
Arithmetic on the company’s own inputs | $25,000 + $8,500 × 60 months = $535,000 | $200,000 + $20,000 + $36,000 × 5 years = $400,000 | (company-claimed inputs) |
Deck illustration, May 2026 | About $500,000 over 5 years; about $100,000 per year; payback described as immediate | About $400,000; payback about 1.1 years | (company-claimed) |
The deck’s rounded $500,000 and the S-4 figure of $535,000 differ, and the company’s own inputs reproduce $535,000. McKinsey puts a typical humanoid bill of materials at about $30,000 to $150,000 per unit (McKinsey, 17 April 2026), which is context for the robot, not for integration. The sources reviewed do not itemise what either deployment fee covers. Acquisition-model crossover analysis sits in RaaS vs Leasing vs Buying Humanoid Robots.
Payback claims set against a second-hand ROI report
An Agility executive stated payback of less than 1 year under RaaS on the 24 June 2026 webcast, and the subscription is described as priced at a discount to the customer’s fully burdened human labour rate (company-claimed). An illustrative comparison of about $100,000 per year for a robot against roughly $200,000 per year of labour is described by BERI as a vendor projection (company-claimed; via secondary press). Set against these, BERI, relaying FreightWaves, reports that GXO’s chief executive said on the 5 August 2026 earnings call that return on investment (ROI) on humanoids had not yet been achieved (reported second-hand, original not retrieved). That statement is not quoted here until the original is checked. Payback figures are inputs to a negotiation, not results.
Integration exposure worksheet
No independent integration cost benchmark was found; values are left blank because nothing public anchors them. The $25,000 and $20,000 deployment fees are the only disclosed anchors (company-claimed). Use the worksheet to make each bidder price, or disclaim, every line. The ROI Model Library carries the sector payback models, and The Hidden Integration Bill covers the cost line items in depth.
Integration exposure worksheet (V9). Buyer-input cells are blank by design; Physical AI Journal does not estimate them.
Cost category | Party that could bear it | Buyer input (amount) | Basis of quote |
Site survey and readiness works | Buyer, integrator or manufacturer | ________ | ________ |
Workflow and system integration (WMS, MES) | Integrator or manufacturer software | ________ | ________ |
Commissioning and acceptance testing | Integrator or manufacturer | ________ | ________ |
Operator and maintenance training | Manufacturer or integrator | ________ | ________ |
Fleet software and monitoring subscription | Manufacturer or software provider | ________ | ________ |
Maintenance, spares and field service | Manufacturer or integrator | ________ | ________ |
Re-integration when the hardware generation changes | To be assigned | ________ | ________ |
Insurance and risk of loss on premises | To be assigned | ________ | ________ |
Data and configuration transfer on exit | To be assigned | ________ | ________ |
Decision implication: require every line to be priced or expressly excluded, and treat any bundled statement that integration is included as unpriced until itemised.
5. Named Deployment Case Studies
Direct answer. In the 4 named deployments, the humanoid maker holds the commercial relationship under vendor-direct RaaS, a captive-parent arrangement or undisclosed terms, while the customer or logistics operator runs the site. No case names a third-party integrator, and none discloses how liability, uptime or acceptance is allocated. Outcome figures are company-claimed. For Schaeffler and Humanoid (UK), no outcome data has been published.
The 4 cases give 4 different structures to compare against the matrix in Section 7: vendor-direct RaaS, a captive parent, undisclosed terms and an announced RaaS agreement, each told through who owned what, as reported. Gaps in disclosure are not a criticism of the announcements; they mark terms a buyer must negotiate privately.
GXO and Agility: vendor-direct RaaS at a logistics operator’s site
At Flowery Branch, Georgia, Digit was reported in June 2024 to pick totes from 6 River Systems Chuck mobile robots and place them on conveyors, orchestrated through Arc; the end client at the site is Spanx and GXO is the operator (company-claimed; via secondary press). Agility is paid under a multi-year RaaS agreement signed on 27 June 2024, which the companies describe as the first RaaS humanoid deployment (company-claimed). No separate integrator is named, and price, term, unit count and liability allocation are not disclosed. The claimed outcome is more than 100,000 totes moved (DC Velocity, 26 October 2026; company-claimed; via secondary press), against over 125,000 on the Agility webcast (company-claimed). The 26 October 2026 expansion starts in early 2027 on Digit 4, moving to Digit 5 once available; the added robot count is not disclosed (company-claimed). Sources conflict on the new sites: Agility’s release names Statesville, North Carolina and a Tennessee site, while The Robot Report, as relayed by BERI, places Statesville in Tennessee and adds Lebanon, Indiana. This is a stated expansion decision, not a delivered fleet, and it is the strongest pilot-to-expansion evidence in the research.
BMW and Figure: a pilot with customer-reported metrics
At Plant Spartanburg, South Carolina, BMW ran Figure 02 in the body shop removing and positioning sheet-metal parts for welding. BMW reports that over 10 months it supported more than 30,000 BMW X3, ran 10-hour shifts Monday to Friday, moved more than 90,000 components, took about 1.2 million steps and logged roughly 1,250 operating hours (company-claimed, customer-side). Figure, The Robot Report and Humanoid Guide give the pilot as 11 months (company-claimed; via secondary press); the difference is not reconciled. The Robot Report reported on 29 June 2026 that BMW will deploy Figure 03 for logistics sequencing; unit count and terms are not disclosed (company-claimed; via secondary press). Humanoid Guide reports that Figure 02 units were retired (low-reliability source). Per The Robot Report, Figure supplies and develops the robots and BMW runs the deployment in its plant; that is not confirmed as a contract allocation. Integrator, price, lease, term, cost, savings and payback were not found. Figure 03 is confirmed only as an announcement.
Hyundai and Boston Dynamics: a captive-parent structure at the training stage
Hyundai Motor Group is the majority shareholder of Boston Dynamics (company-claimed), so deployment is captive-parent and no third-party integrator is named. The Robotics Metaplant Application Center (RMAC) at Metaplant America near Savannah, Georgia, was announced on 21 September 2026 as a testbed for integrating Atlas into vehicle manufacturing (company-claimed). Its first phase is operational and Atlas is training on logistics and parts sequencing (company-claimed); component assembly is planned by 2030 and a building about 10 times larger in 2027 (announced target). Hyundai plans 25,000 Atlas units across Hyundai and Kia plants over the next few years and a US facility producing 30,000 robots per year (announced target); the year for that facility is not consistently stated, and The Robot Report cites 2028 as industry context. Outcome data on cycle time, accuracy, uptime and autonomy were not found. Status: training and announcement, not production deployment.
Schaeffler and Humanoid (UK): announcement only
Humanoid (UK) and Schaeffler announced a binding phased deployment and supply agreement on 13 May 2026 (RoboticsTomorrow; company-claimed); AI Business dates it 18 May 2026, and the 2 outlets also differ on unit wording, with AI Business reporting thousands of robots and RoboticsTomorrow reporting 4-digit wheeled units by 2032 (announced target; not a confirmed order). The robot is the wheeled HMND 01 under RaaS including fleet software, maintenance and 24/7 support, with Bosch as contract manufacturer for Europe (company-claimed). The initial phase runs from December 2026 to June 2027: box handling in live production at Herzogenaurach, and at Schweinfurt a 3-month integration and demonstration followed by 3 months of validation (announced target). AI Business notes that the reporting does not say who maintains or operates the robots after deployment. Financial terms and robot counts are not disclosed. Only the announcement is confirmed; no outcome data has been published.
Case accountability map (V10 and V11 combined). “Not disclosed” means the sources reviewed do not state it. All entries are company-claimed.
Question | GXO / Agility | BMW / Figure | Hyundai / Boston Dynamics | Schaeffler / Humanoid (UK) |
Structure | Vendor-direct RaaS | Not disclosed; BMW runs the deployment per trade press | Captive parent | Announced RaaS |
Third-party integrator named | No | No | No | No |
Robot ownership | Agility retains under RaaS | Not disclosed | Not disclosed; captive parent | Not disclosed beyond RaaS |
Site operation | GXO | BMW | Hyundai plants (inferred) | Not stated |
Maintenance after deployment | Not disclosed; Agility’s RaaS description includes maintenance | Not disclosed | Not disclosed | In RaaS scope; operator not stated |
Liability and acceptance terms | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
Outcome evidence | Company-claimed totes | Company-claimed, customer-side metrics | None | None |
Status | Expansion announced | Figure 02 reported retired; Figure 03 announced | Training stage | Deployment from December 2026 |
The pattern is consistent: vendor-direct RaaS, captive-parent or undisclosed structures, 1 announcement-only agreement, and no integrator named.
6. Friction, Risk & Unresolved Issues
Direct answer. The main unresolved risks are contractual: no public humanoid contract discloses service levels, liability caps or indemnities; hardware generations change inside programmes; counterparties can fail; unit claims are often untraceable; and humanoid-specific standards are not final. Regulation is reported here as dated status only. Each unresolved item should be assigned by contract because no standard allocates it. Items marked not found were searched for and not located.
Friction register (V12). Only frictions that bear on accountability are listed.
Issue | Evidence | Procurement implication |
Hardware-generation churn | Figure 02 units reported retired after the Spartanburg pilot (low-reliability source); GXO’s new sites start on Digit 4 in early 2027, then move to Digit 5 (company-claimed) | Assign who bears re-integration work and cost |
Counterparty fragility | K-Scale Labs disbanded in November 2025 (low-reliability source: secondary compilation); Agility reported 2025 net sales of $1.8 million against an operating loss of $140 million (company-claimed; via secondary press) | Test financial strength; require continuity and step-in rights |
Unverifiable unit claims | Claims of over 50,000 Optimus units and over 10,000 Figure deployments have no traceable company disclosure (low-reliability source); Tesla’s Q3 2026 production release to the SEC contains no Optimus information | Require counts in writing, with site names |
Contract opacity | No humanoid-specific service-level, liability, indemnity, insurance, warranty, update-rights or data-ownership terms found from any primary source; Foley Hoag’s 15 October 2026 RaaS commentary does not mention humanoids (secondary legal commentary) | No market-standard terms exist; the buyer must propose them |
Pilot-to-scale gap | McKinsey (17 April 2026) reports that none of 3 industrialisation triggers, including predictable multi-year volumes, is fully in place | Do not build prices or schedules on assumed volume |
Country-of-origin screen | FCC action of 28–29 July 2026; effective date and exemptions not stated in the AP report | Ask each bidder for country of origin and sourcing plan; counsel confirms |
Secondary legal commentary raises, but does not settle, who becomes the manufacturer of an integrated cell. Unchained Robotics (19 August 2026) states that the answer depends on project structure and the division of roles, and that a party making a substantial modification is treated as a manufacturer (secondary legal commentary; vendor blog; EUR-Lex not checked). No source tested whether a software update on a humanoid shifts that role. WTW (26 August 2026) lists manufacturer, software developer, operator and maintenance provider as possible loci of liability (secondary legal commentary).
Standards and regulation status (V13). Dated facts only.
Item | Status as reported | Source |
ISO 10218-1:2025, industrial robots, Part 1 | Published February 2025, Edition 3, ISO/TC 299 | ISO, accessed 2 November 2026 |
ISO 10218-2:2025, robot applications and cells | Date and status not found; trade press says Part 2 is aimed at system integrators | Robotics 24/7, 18 February 2025 |
ISO/CD 25785-1, dynamically stable industrial mobile robots | Committee Draft consultation; progress beyond draft not found | ISO, accessed 2 November 2026 |
Final humanoid-specific ISO, IEC, IEEE or A3 standard | Not found; IEEE Humanoid Study Group expects 18 to 36 months for standards development | The Robot Report, 25 September 2025 |
EU Machinery Regulation (EU) 2023/1230 | Applies from 20 January 2027; EUR-Lex text not found | Secondary legal commentary |
EU AI Act timing (Digital Omnibus on AI) | In force 27 July 2026; Annex III high-risk moved to 2 December 2027, Annex I to 2 August 2028 | Lewis Silkin, 27 July 2026 (secondary legal commentary) |
EU Product Liability Directive (EU) 2024/2853 | National implementation due 9 December 2026 | SKW Schwarz, 21 January 2026 (secondary legal commentary) |
Not found in the sources reviewed: any named third-party integrator in a humanoid deployment case; any OEM-integrator alliance or channel programme; humanoid-specific contract terms from a primary source; per-hour pricing; lease structures distinct from RaaS; integration costs; independently measured uptime or throughput; teleoperation share; BMW/Figure cost, savings or payback; UK humanoid-specific guidance and an OSHA position; and named cancellations or scaled-back pilots. Absence in a limited search is not evidence that none occurred.
Decision implication: every row above is a clause to write, not a risk to assume away.
7. Competitive / Platform Comparison
Direct answer. OEM-led integration suits a single-platform deployment where the buyer accepts vendor dependency for speed. An independent integrator suits multi-vendor or multi-site estates that need neutrality, but no independent humanoid integrator has a disclosed outcome yet. A consortium needs a named lead party. In-house integration needs engineering capacity the buyer already holds. No public humanoid evidence was found for the consortium or in-house models. This is Physical AI Journal analysis.
No public evidence ranks the 4 integration models, so the comparison runs in 3 steps: models, then the responsibility matrix, then platform pathways. Integration-pathway fields feed the Platform Index.
Integration-model comparison (V14). Physical AI Journal analysis; ratings are qualitative and rest on the public evidence in Sections 2 to 6.
Criterion | OEM-led | Independent integrator | Consortium with lead party | In-house |
Accountable counterparty | One, if the contract assigns it (a design objective) | One, if it accepts acceptance and performance duties | One only if a lead party is named | The buyer |
Integration depth | Through the maker’s own software layer (Arc, Orbit; company-claimed) | Across WMS, MES and site systems, as scoped | Combined, with interface seams | Highest control; limited by buyer capacity |
Commercial transparency | Pricing disclosed for Agility only (company-claimed) | Fees not disclosed; no benchmark | Several fee layers | Internal cost visibility |
Vendor dependency | High | Maker dependency remains; integrator dependency added | Spread across parties | Maker only |
Scalability across sites | Follows the maker’s rollout | Portable if hardware-agnostic (Robo Inc. so describes itself; company-claimed) | Coordination load rises per site | Limited by team size |
Buyer leverage | Depends on volume and exclusivity | Higher when scored competitively | Mixed | Highest over integrators; maker terms still apply |
Public humanoid evidence | Structures disclosed (Section 5) | None with outcomes | None found | None found |
No model dominates on the public evidence. OEM-led integration is the only route with disclosed structures, but it concentrates dependency in one party and ties buyer leverage to volume and exclusivity. An independent integrator adds neutrality and system-integration depth, but no reviewed case supplies outcomes, so the buyer carries the qualification burden. A consortium works only when one lead party holds the contract. In-house integration suits a buyer with engineering capacity on a single platform, and it leaves the maker’s terms unchanged. Match the model to the estate, then test it with the scorecard in Section 8.
Responsibility matrix (V15). Physical AI Journal proposed default allocation for an independent-integrator appointment, not drawn from any disclosed contract.
A = accountable, R = responsible, C = consulted, I = informed.
The software provider may be the manufacturer.
Deliverable | OEM | Integrator | Software provider | Operator |
Hardware warranty | A/R | I | C | I |
Workflow and system integration (WMS, MES) | C | A | R | C |
Commissioning | R | A | C | C |
Site readiness | I | C | I | A/R |
Operational acceptance | C | R | C | A |
Ongoing maintenance | R | A | C | I |
Software updates | C | C | A/R | C |
Regulatory conformity documentation | † | † | † | † |
Data and configuration ownership | C | R | C | A |
† The contract must name the holder. Secondary legal commentary says a buyer of a complete robot cell would normally expect system-level conformity documents; counsel decides. This report gives no view on any robot’s conformity.
Platform integration pathways (V16). Manufacturer-claimed pathway only; no independent verification of any integration pathway was found.
Platform | Pathway (manufacturer-claimed) | Independently verified |
Agility Digit | Arc software with RaaS or Ownership; Arc integrates Digit into facilities and workflows (company-claimed) | None |
Boston Dynamics Atlas | Orbit software connects to MES, WMS and other industrial systems; captive parent Hyundai (company-claimed) | None |
Humanoid (UK) HMND 01 | RaaS with fleet software, maintenance and 24/7 support; Bosch contract manufacturing (company-claimed) | None |
UBTech Walker S2 | Direct supply the company calls a “standardized, replicable turnkey solution” (company-claimed) | None |
No OEM-integrator alliance or channel programme with a date and source was found for Agility, Apptronik, Figure, Unitree or 1X. Unitree’s pathway was not established in the research and is left out rather than inferred.
Decision implication: pick the model that matches your estate, then use the matrix to write down who holds each row before negotiating price.
8. Strategic Recommendations
Direct answer. Define the outcome first, then choose the integration model, score candidates on evidence, tie payment to accepted and sustained results, and set the conditions under which appointment is deferred. A buyer that cannot obtain classified deployment evidence, a written responsibility schedule and acceptance tests before signature should defer. The decision framework, scorecard and checklists below are Physical AI Journal proposals; thresholds are set by the buyer, and counsel should review contract terms.
Appointment decision framework
Defer unless the evidence meets a buyer-set threshold; no reviewed case supplies the evidence an independent appointment needs today.
Appointment decision framework (V17). Deferral conditions are buyer-set criteria.
Outcome | Choose when | Evidence required before signature |
Appoint OEM-direct | A single platform, 1 or 2 sites, integration limited to the maker’s software layer | Written responsibility schedule; acceptance tests; named holder of conformity documentation; payment tied to acceptance |
Appoint an independent integrator | Multi-vendor or multi-site estate, or deep WMS and MES work | Site-named references with operating results, or a buyer-set minimum scorecard score; no reviewed case supplies this today, so set the bar higher |
Structure a consortium | No single party covers the scope | One lead party holding the single contract, with back-to-back responsibilities |
Defer | Evidence is below the buyer’s threshold, the hardware generation is changing, no baseline exists, or no party will hold conformity documentation | A written list of the conditions that would end the deferral |
Weighted qualification scorecard
Verified deployment evidence carries the largest proposed weight, 25 of 100, and the 6 criteria total 100.
Qualification scorecard (V18). Weights are Physical AI Journal proposals and total 100; adjust them to the buyer’s risk profile.
Criterion | Weight | Evidence to request | Red flag |
Verified deployment evidence | 25 | Named sites, dates, operating hours, intervention logs | Announcements only, or totals without site or hours |
Acceptance and accountability willingness | 20 | Draft responsibility schedule and acceptance tests | Refusal to tie any payment to accepted results |
Commercial and financial strength | 15 | Audited accounts, funding position, parent support | Revenue resting on announced backlog; Agility’s deck itself says backlog headlines are not current-period revenue (company-claimed) |
Service coverage and continuity | 15 | Field-service coverage near each site, spares, response commitments, step-in rights | Single-region support for a multi-site plan |
Systems integration capability | 15 | WMS and MES references, disclosed interfaces | Integration scope that relies on vendor software alone |
Data, IP and portability terms | 10 | Ownership of configuration, code and data; export format | Data used for model training with no buyer control |
Score each criterion from 0 to 5 against the evidence supplied, multiply by its weight and sum. The buyer sets the minimum total and any criterion that works as a pass-fail gate; Physical AI Journal proposes no threshold. Give credit only for evidence the bidder can document: announcements, press releases and company-claimed totals earn credit for the specific site, date and hours disclosed, and no more. Re-score when the hardware generation changes.
Contract provision checklist
The checklist covers 7 provisions, limited to appointment architecture.
Contract provisions limited to appointment architecture (V19). The liability framework in Humanoid Deployment Risk.
Provision | Buyer ask |
Scope and responsibility schedule | Every row of the matrix assigned to a named party |
Acceptance tests and observation period | The 6 indicators with buyer-set thresholds and window |
Payment release | Tied to accepted, sustained results, not to demonstration |
Change control | Written process for hardware-generation and software changes, with cost allocation |
Service levels and remedies | Task-level measures, not hardware availability alone; remedies such as service credits and fee reductions |
Data and configuration ownership | Export, return and use limits stated |
Step-in and continuity rights | Rights if the integrator or manufacturer fails |
Foley Hoag’s RaaS commentary gives an uptime example of 95% of scheduled operating hours with service credits, fee reductions and termination for chronic underperformance (secondary legal commentary; not humanoid-specific). Law-firm guidance treats uptime service levels as hardware-availability measures, and no source reported humanoid-specific service-level metrics, so the task-level indicators in Section 3 are the better contract anchor.
Buyer action list: first 90 days
The first 90 days run in 3 windows: define and map, score and negotiate, then decide.
90-day action list (V20).
Window | Actions |
Days 1–30 | Define the outcome and capture the baseline; map every deliverable to a party with the matrix; request written responsibility schedules from the manufacturer and any integrator; ask counsel who holds conformity documentation |
Days 31–60 | Score candidates on site-named evidence; draft acceptance tests with Operations; negotiate payment triggers and the integration exposure worksheet |
Days 61–90 | Decide to appoint or defer; sign with change control and step-in rights; set the observation period and reporting cadence |
Operations Leader validation list
Before sign-off, Operations should confirm that baseline data exist for each task; that a named owner holds site readiness; that shift handover and intervention procedures are written; that maintenance access windows fit the production calendar; that the observation period fits the production schedule; and that a plan exists for the hardware-generation change.
Decision for the Automation/Procurement Buyer: before signature, obtain from each candidate a written responsibility schedule, site-named deployment evidence and acceptance tests with buyer-set thresholds. If all 3 are in hand, appoint the model that scores highest on the scorecard and tie payment to accepted, sustained results; if any is missing, defer and record the conditions for revisiting.
Continue reading part ii - https://www.physicalaijournal.org/post/who-owns-the-humanoid-deployment-part-ii