Humanoid Robot Vendor Due Diligence 2026: 20 Questions Before You Sign
- Physical AI Team
- 5 days ago
- 26 min read
Physical AI Journal Research Team — Independent analysis by Sekason Research Limited, United Kingdom
How industrial buyers should evaluate deployment evidence, procurement economics, integration risk, vendor resilience and contractual commitments before committing to humanoid robotics.
Executive Intelligence Synthesis
What is the most important question to ask a humanoid robot vendor?
Ask for independently corroborated evidence of repeatable, measurable performance in a comparable operating environment — not a controlled demonstration. The critical distinction for 2026 procurement is between a vendor that can demonstrate a robot and a vendor that has demonstrated sustained operation inside a real industrial site, supported a named customer for a meaningful number of operating hours, and produced verifiable task-level outcomes that hold under normal production variability.
Industrial buyers evaluating humanoid robot procurement in 2026 must choose between vendors with verifiable operating evidence and vendors with demonstrations and future promises — only the former category warrants capital commitment. The evidence base is narrower than the market narrative suggests: Interact Analysis estimates that only around 10% of the more than 20,000 humanoid units produced globally in 2025 were deployed in real-world applications.
This report delivers a structured due diligence framework — Green / Amber / Red — that tells operations leaders and procurement buyers whether to commit to a production contract, proceed to a structured pilot only, or wait.
Five strategic signals, drawn from the strongest public evidence available as of August 2026, frame the buyer's evaluation criteria.
Signal 1 — Deployment evidence is now more valuable than demonstrations.
BMW Group disclosed that Figure 02 supported production of more than 30,000 BMW X3 vehicles over approximately 10 months, accumulated approximately 1,250 operating hours, moved more than 90,000 components, and covered around 1.2 million steps in its Spartanburg, South Carolina facility (customer-disclosed). That level of publicly reported operating activity is currently rare. It sets a benchmark that buyers should use to interrogate every other vendor's evidence claims — and most claims will not survive the comparison.
Signal 2 — Production volume is not deployment maturity.
Interact Analysis estimates that more than 20,000 humanoid units were produced globally in 2025 — approximately ten times 2024 production — but only around 10% of those units were deployed in real-world applications (verified, analyst research). A vendor producing thousands of units at a factory is not the same as a vendor operating thousands of units at customer sites. Procurement must make this distinction explicit in every RFQ.
Signal 3 — The real purchase is a robot ecosystem, not hardware.
Every major commercial deployment reviewed in this research involved the robot, software, fleet management infrastructure, integration engineering, teleoperation or human intervention protocols, and ongoing support. Agility Robotics offers its Agility Arc cloud platform alongside Digit (company-claimed). Figure AI integrates with customer production IT. Apptronik has built a dedicated real-world data collection facility to accelerate the transition from pilot to production (verified, Reuters, 30 June 2026). Buyers who price only the robot will build the wrong business case.
Signal 4 — Enterprise pricing and five-year TCO remain poorly disclosed.
No authoritative public enterprise purchase price was identified in this research for Figure 03, Digit, Apollo, or Atlas. Publicly listed pricing exists only for lower-cost platforms — Unitree G1 from $13,500 (company-claimed) and 1X NEO from $20,000 (company-claimed) — and neither is a direct industrial benchmark for enterprise procurement. Every buyer is currently negotiating in an information vacuum. That is not a temporary market condition; it is an active commercial risk.
Signal 5 — The decision is Buy / Pilot / Wait, not Buy / Don't Buy.
The International Federation of Robotics (IFR) characterised the market in June 2026 as one where reliability and efficiency still need to be proven, with most applications still trials or prototypes (verified, IFR Executive Roundtable). The Green / Amber / Red framework this report delivers reflects that commercial reality: some vendors warrant commitment, most warrant a structured pilot, and some do not yet provide sufficient evidence to justify either.
Signal | What changed | Procurement implication |
Deployment evidence | Sustained production pilots now publicly documented | Demand operating-hour evidence, not just demonstrations |
Production volume | Exceeds real-world deployment volume by ~10× | Treat production capacity as a separate question from delivery capability |
Autonomy level | Human intervention remains embedded in commercial deployments | Measure intervention rates; do not accept unqualified autonomy claims |
Pricing | Enterprise TCO remains largely undisclosed | Build scenario models from vendor quotes, not market averages |
Ecosystem | Robot + software + service + data is the actual product | Contract for the full system, not the hardware unit |
Platform and Market Landscape
Which humanoid robot vendors are relevant for industrial procurement in 2026?
The relevant vendor universe for industrial buyers in 2026 includes platforms with at least one named industrial customer, a commercial agreement, or a credible production programme with disclosed timelines. Announced prototypes without customer validation do not qualify. On that basis, Figure AI, Agility Robotics, Apptronik, Boston Dynamics / Hyundai, and Humanoid (the company, with its HMND 01 platform) represent the platforms with the most substantiated industrial procurement relevance. Unitree and 1X are relevant context but are not positioned for enterprise industrial procurement at equivalent maturity.
By August 2026, the humanoid vendor universe divides into three tiers: a small number with published industrial deployment evidence that warrants procurement scrutiny, a larger group that has raised substantial capital and is scaling manufacturing, and a significant portion whose announced units remain in development, internal testing, or research use. This distribution is the starting point for vendor evaluation — the tier a vendor occupies determines the procurement questions a buyer should ask.
Figure AI has the most extensively publicly documented industrial deployment record of the vendors reviewed. The Figure 02 programme at BMW Spartanburg currently provides the strongest single public evidence point available to procurement buyers in this research period. Figure 03 arrived at BMW's Hall 52 on 30 June 2026 for a new logistics sequencing workflow (company-claimed / customer confirmation). Figure reports its BotQ facility produced more than 350 Figure 03 robots and increased production rate from one robot per day to one robot per hour in under 120 days (company-claimed). These production figures should be verified by any buyer before treating them as contractual capacity commitments.
Agility Robotics is the only vendor to have announced a multi-year commercial Robots-as-a-Service (RaaS) agreement with a named logistics customer — GXO Logistics — as of the date of this research. Agility reports active deployments with Schaeffler, GXO, Toyota Motor Manufacturing Canada, and Mercado Libre, and more than $300 million in multi-year Digit v5 orders subject to contractual milestones (company-claimed, 16 July 2026). Buyers should note that "orders subject to contractual milestones" is not the same as shipped robots operating at customer sites.
Apptronik closed a $520 million Series A-X extension on 11 February 2026, bringing its Series A financing to more than $935 million and total capital raised to nearly $1 billion (verified, Reuters). The round valued Apptronik at approximately $5 billion (verified, Reuters). Its Apollo platform is deployed in a testing and production-environment learning programme at Mercedes-Benz's Berlin-Marienfelde facility. Apollo 2 was unveiled on 30 June 2026, alongside a dedicated robot training and data collection facility (verified, Reuters).
Boston Dynamics / Hyundai Motor Group describes its production Atlas as having 56 degrees of freedom, 360-degree vision, a maximum payload of 50 kg, autonomous battery replacement, and water/dust resistance (company-claimed). Hyundai plans to deploy Atlas at its manufacturing sites, with a planned rollout at its Georgia operation beginning in 2028 (verified, Reuters, 6 January 2026). Broad production deployment of Atlas remains a future programme, not a current installed base.
Unitree priced its Shanghai IPO at 150.80 yuan per share on 10 August 2026, seeking approximately 6.1 billion yuan (~$904 million), becoming the first mainland-listed humanoid robot manufacturer (verified, Reuters). 2025 revenue was approximately 1.7 billion yuan, with more than 40% of sales from overseas (verified, Reuters). Unitree's G1 carries a published purchase price of $13,500 (company-claimed) — important evidence that transparent humanoid pricing is commercially possible, but not a direct industrial automation benchmark.
Vendor | Platform | Industrial evidence | Commercial model | Public pricing | Evidence status |
Figure AI | F.02 / F.03 | BMW Spartanburg deployment | Quote / pilot | Undisclosed | Customer-disclosed (F.02); company-claimed (F.03) |
Agility Robotics | Digit | GXO commercial RaaS | RaaS / commercial | Undisclosed | Customer-disclosed |
Apptronik | Apollo / Apollo 2 | Mercedes-Benz production testing | Commercial / pilot | Undisclosed | Verified (Reuters) |
Boston Dynamics / Hyundai | Atlas | Planned 2028 deployment | Enterprise | Undisclosed | COMPANY-CLAIMED (specs); VERIFIED — Reuters (2028 deployment plan) |
Humanoid (HMND 01) | HMND 01 | Schaeffler planned 2026–27 | Commercial | Undisclosed | Verified (Reuters) |
Unitree | G1 | Research / development | Purchase | $13,500 (company-claimed) | Company-claimed |
1X | NEO | Home / consumer | Purchase / subscription | $20,000 / $499/month (company-claimed) | Company-claimed |
This table is a procurement landscape, not a ranking. All vendor-originated figures are company-claimed unless otherwise noted.
Deployment Evidence — Reported Outcomes, Not Just Announcements
How should buyers evaluate evidence from a humanoid robot vendor?
Buyers should classify every vendor evidence claim against a seven-stage deployment maturity ladder:
Announcement → Demonstration → Named Pilot → Paid Pilot → Operational Deployment → Sustained Operation → Measured Customer Outcome.
The majority of humanoid vendor claims as of 2026 sit in the lower half of this ladder. Only a small number of deployments — led by BMW / Figure AI and GXO / Agility Robotics — have reached operational deployment with disclosed task-level evidence. Independent verification of customer-disclosed outcomes remains limited across all vendors.
The evidence problem in humanoid procurement is structural, not temporary. A vendor that issues a press release about a "deployment" may be describing an announcement, a controlled demonstration, a short-duration test, a paid commercial contract, or a robot that has been operating autonomously for thousands of hours. The buyer has no automatic mechanism to distinguish between these — unless they demand it.
The Evidence Ladder defines the minimum acceptable classification for procurement purposes:
Announcement — a vendor or customer has publicly stated an intent to deploy. No robot is yet operating at a site.
Demonstration — a vendor has shown the robot performing a task in a controlled environment. Performance may not reflect real production variability.
Named Pilot — a named customer has allowed a vendor to test a robot on site. The pilot may be unpaid, short-duration, or supervised. Evidence may be disclosed only by the vendor.
Paid Pilot — a commercial agreement exists. The customer is paying for the pilot. Task performance acceptance criteria may or may not be defined.
Operational Deployment — the robot is performing production work. Operating hours and task cycles are accumulating. Customer is receiving economic value.
Sustained Operation — the robot has operated continuously for a meaningful period across full production shifts, without the pilot being terminated or reset.
Measured Customer Outcome — the customer has disclosed specific operating data, task-level results, labour impact, or economic outcomes attributable to the deployment.
Against this ladder, the BMW / Figure 02 deployment occupies the highest publicly documented position for any enterprise humanoid at the time of this research. BMW customer-disclosed 1,250 operating hours, more than 90,000 components moved, and production support for more than 30,000 BMW X3 vehicles. Crucially, the evidence was disclosed by the customer, not only by the vendor — which places it above the "company-claimed" classification used for vendor-only announcements, while remaining customer-reported rather than independently audited.
GXO / Digit has reached operational deployment with a publicly disclosed task-volume milestone — more than 100,000 totes moved (company-claimed, Agility Robotics, November 2025). The commercial structure is a multi-year RaaS agreement signed 27 June 2024 (verified, GXO). What the public record does not provide is independent breakdown of uptime, labour displacement, operating cost, or buyer ROI. That gap is analytically important: a task volume milestone is not the same as a measured commercial outcome.
Mercedes-Benz / Apollo has reached the Named Pilot / Paid Pilot stage in the ladder. Mercedes-Benz has confirmed that Apollo is operating in its production environment at Berlin-Marienfelde, collecting data, receiving knowledge transfer from employees through teleoperation and augmented reality, and moving toward autonomous capability — not operating fully autonomously as of the date of the research. A production-environment learning deployment and a finished automation product are different stages.
The BMW Leipzig / AEON (Hexagon) programme, announced 27 February 2026, represents a new-format pilot: structured assessment under BMW's formal evaluation sequence (theoretical → laboratory → test deployment → pilot) (verified, BMW Group). The programme's specific use cases were not publicly detailed in the source reviewed. It illustrates how a sophisticated buyer builds the evaluation sequence into the vendor relationship from the outset, rather than accepting a demonstration as a substitute for structured procurement validation.
The Schaeffler / Humanoid planned deployment — announced 13 May 2026 and reported by Reuters — targets 1,000–2,000 robots across Schaeffler's global manufacturing sites by 2032, with initial deployment scheduled between December 2026 and June 2027 at Herzogenaurach and Schweinfurt. This is a planned deployment commitment, not evidence of robots already operating. Buyers should note that Reuters reported contract value and precise deployment numbers were not disclosed by the company.
Humanoid Evidence Ladder — Procurement Classification
Stage | Description | Example (as of Aug 2026) |
Announcement | Intent declared, no robot on site | Many vendor press releases |
Demonstration | Controlled task performance | Most vendor showcases |
Named Pilot | Named customer, on-site test | Various undisclosed trials |
Paid Pilot | Commercial agreement, test conditions | Multiple ongoing |
Operational Deployment | Robot performing production work | GXO / Digit; BMW / Figure 02 |
Sustained Operation | Multi-month continuous production activity | BMW / Figure 02 (~10 months) |
Measured Customer Outcome | Customer-disclosed operating data | BMW / Figure 02 (customer-disclosed) |
Economics and ROI Analysis
How do you calculate the ROI of a humanoid robot?
ROI for a humanoid robot deployment cannot be calculated from the hardware price or RaaS fee alone. The correct approach is a task-specific total cost of ownership (TCO) model that captures every cost category — hardware or lease, integration, site preparation, software, connectivity, training, maintenance, spare parts, downtime, human intervention, remote support, insurance, and end-of-life — and compares the full cost against the quantified value of productive output. As of August 2026, no independently verified five-year TCO for any major humanoid deployment is publicly available. That absence is itself a procurement finding.
Enterprise pricing for the platforms most relevant to industrial buyers — Figure 03, Digit, Apollo, and Atlas — is not publicly disclosed. This is not a minor data gap; it is a structural commercial condition that puts buyers at a systematic disadvantage in the first procurement cycle. The report's central commercial recommendation is therefore to build the TCO model from the actual vendor quote, not from any market average or reported headline figure.
The complete TCO architecture for humanoid procurement includes:
Hardware / lease cost — the robot unit price or RaaS monthly fee. For RaaS, calculate the full contract duration cost before comparing with purchase.
Integration cost — APIs, interfaces, conveyor modifications, fixture changes, safety enclosures, floor marking, IT integration, ERP/WMS/MES connectivity. BMW's Figure 02 deployment required integration through standardised interfaces in BMW's Smart Robotics ecosystem — integration was not trivial. Who pays for this in the proposed contract?
Site preparation — floor load capacity, power supply, lighting, network connectivity, safety zoning, access restrictions. These costs are frequently excluded from vendor pricing and often underestimated by buyers.
Software licensing — fleet management, teleoperation platforms, AI model licensing, update fees, API access, usage-based pricing. Agility's Arc platform (company-claimed) and Figure's production software ecosystem (company-claimed) are described by their respective vendors as integral components of the deployment, not separate optional purchases.
Connectivity — cellular, Wi-Fi, private 5G for real-time control and monitoring. Costs vary significantly by facility.
Training — initial operator training, ongoing workforce adaptation, safety certification for personnel operating alongside the robot. Mercedes-Benz's Apollo programme explicitly included knowledge transfer from employees through teleoperation before the robot moved toward autonomous operation. That time has a cost.
Maintenance — scheduled preventive maintenance, firmware updates, calibration. Who delivers this? What is the response time? What is the contractual SLA?
Spare parts — availability, lead times, pricing. For a vendor with limited installed base, spare parts supply chain may be unproven at scale.
Downtime cost — the production value lost while the robot is offline for maintenance, repair, software updates, or failure. For a robot performing a critical production task, downtime cost can exceed maintenance cost.
Human intervention cost — the labour cost of employees required to supervise, correct, or remotely operate the robot during its current intervention rate. Agility's CEO acknowledged in a published interview that fully AI-controlled operation is not the current approach for its heavy commercial robot. Buyers should quantify the intervention rate and cost it explicitly.
Remote support — vendor-provided remote monitoring, troubleshooting, and software updates. What is included? What is charged separately?
Insurance and liability — additional product liability coverage, operational risk assessment updates, and employer obligations for human-robot shared workspaces.
End-of-life and replacement — what happens when the robot reaches end of commercial life or the vendor launches a successor model? What are the upgrade path terms, trade-in rights, and software continuity provisions?
The absence of verified five-year TCO data is itself a finding.
This research did not identify independently verified five-year TCO figures for any of the major industrial humanoid deployments. Every buyer must therefore build their own model — and every model will rely heavily on vendor-provided assumptions for at least the first procurement cycle. That makes the quality of vendor disclosure one of the most important procurement quality signals available.
Three-Scenario TCO Framework — Base / Downside / Upside
Variable | Downside | Base | Upside |
Productive hours per shift | Lower than specification | Vendor estimate | Specification achieved |
Human intervention rate | High (>20% of tasks) | Moderate | Low (<5% of tasks) |
Unplanned downtime | High | Moderate | Low |
Integration cost | 2× vendor estimate | 1.5× vendor estimate | Vendor estimate |
Maintenance cost | Higher than contracted | Per contract | Lower than contracted |
Task throughput | 70% of target | 90% of target | 100%+ of target |
No standard humanoid ROI figure is provided in this report. The research does not support one. A buyer who accepts a vendor's ROI projection without building a scenario model is accepting a single-point estimate in a market where the evidence base for those estimates remains thin.
Named Deployment Case Studies
Which humanoid robot deployments provide the strongest public evidence for industrial buyers?
As of August 2026, the deployments with the strongest publicly available evidence for industrial buyers are BMW / Figure 02 at Spartanburg (customer-disclosed operating hours, task volume, and production context), GXO / Digit at Flowery Branch, Georgia (commercial RaaS agreement with a disclosed task-volume milestone), and Mercedes-Benz / Apollo at Berlin-Marienfelde (production-environment testing confirmed by Reuters and customer disclosure). Each case provides different procurement lessons. None provides independently audited commercial ROI.
Case Study 1 — BMW Group / Figure 02 | Spartanburg, South Carolina
Sector: Automotive manufacturing | Platform: Figure 02 | Timeline: Approximately 10 months, 2025
BMW Group disclosed that Figure 02 supported production of more than 30,000 BMW X3 vehicles, moved more than 90,000 components, accumulated approximately 1,250 operating hours, and covered approximately 1.2 million steps. The robot operated on ten-hour shifts, Monday to Friday. Its use case was sheet-metal removal and positioning for welding. BMW also reported that laboratory-trained motion sequences transferred into stable shift operation faster than expected, and that integration was implemented through standardised interfaces within BMW's Smart Robotics ecosystem (customer-disclosed / CUSTOMER-REPORTED).
What the evidence confirms: Sustained multi-month operation in a real automotive production environment. Task specificity — a defined, repeatable industrial task, not a general-purpose demonstration.
What the evidence does not confirm: Independent verification of stated figures. Five-year TCO. Labour displacement economics. Uptime benchmarks. Failure rates. Comparative cost versus alternative automation.
Procurement lesson: BMW's prior evaluation sequence — theoretical assessment, laboratory evaluation, test deployment, then pilot — is a model for how sophisticated buyers approach humanoid procurement. The data disclosed is unusually specific for this market. Use it as the baseline for interrogating every other vendor's evidence package.
Evidence classification: CUSTOMER-REPORTED / COMPANY-CLAIMED (Figure AI)
Case Study 2 — GXO Logistics / Digit | Flowery Branch, Georgia
Sector: Logistics / warehousing | Platform: Agility Robotics Digit | Timeline: Pilot from late 2023; commercial RaaS agreement from 27 June 2024
GXO Logistics signed what it described as the industry's first formal commercial deployment of a humanoid robot and the first humanoid RaaS deployment, with Agility Robotics at its SPANX facility in Flowery Branch, Georgia. Digit performed repetitive material-handling tasks — moving totes from cobots and placing them onto conveyors. Agility subsequently reported that Digit moved more than 100,000 totes at the facility (company-claimed, November 2025).
What the evidence confirms: A multi-year commercial agreement with a named logistics customer. A task-volume milestone publicly reported by the vendor. The first commercial humanoid RaaS structure in the public record.
What the evidence does not confirm: Independent breakdown of uptime, labour hours displaced, operating cost, customer ROI, or Digit's performance relative to alternative automation options. The 100,000 totes figure is company-claimed — GXO's own investor communications have not, to this research's knowledge, independently corroborated the specific figure with financial or operational detail.
Procurement lesson: A commercial RaaS agreement is stronger evidence than a pilot MOU, but it is not evidence of economic performance. Buyers should ask for customer-level productivity and cost data, not vendor-reported task volume milestones.
Evidence classification: CUSTOMER-DISCLOSED (GXO commercial agreement) / COMPANY-CLAIMED (Agility task figures)
Case Study 3 — Mercedes-Benz / Apollo | Berlin-Marienfelde, Germany
Sector: Automotive manufacturing | Platform: Apptronik Apollo | Timeline: Testing announced March 2025; ongoing as of June 2026
Mercedes-Benz confirmed it was testing Apollo in a production environment and at its Digital Factory Campus. Initial use cases included transporting components to production lines and performing initial quality checks. Mercedes-Benz also disclosed that employees transferred production knowledge to Apollo through teleoperation and augmented-reality processes — and that the next stage involved teaching Apollo to operate autonomously (customer disclosure / verified, Reuters, 18 March 2025).
What the evidence confirms: A real automotive production environment is being used as a learning and validation stage for Apollo. A strategic relationship exists — Mercedes-Benz has taken a stake in Apptronik (verified, Reuters). Human intervention and knowledge transfer are explicit, structured parts of the deployment methodology.
What the evidence does not confirm: Autonomous production performance. Throughput metrics. Five-year cost economics. Task success rates without human intervention. Timeline to full autonomous operation.
Procurement lesson: This case is analytically valuable precisely because it does not overstate the robot's current status. A production-environment learning deployment is not a finished automation product. Buyers should interrogate every vendor's deployment claims to establish whether the robot is operating autonomously, with teleoperation, or with structured human intervention — and what the commercial economics are at each stage.
Evidence classification: VERIFIED (Reuters / Mercedes-Benz customer disclosure)
Case Study 4 — BMW Group / AEON (Hexagon) | Leipzig, Germany
Sector: Automotive manufacturing | Platform: AEON (Hexagon) | Timeline: Announced 27 February 2026
BMW Group announced its first European humanoid pilot at its Leipzig plant, working with Hexagon's AEON platform.
BMW's Senior Vice President, Production Network, Michael Nikolaides, stated:
"Pilot projects help us to test and further develop the use of Physical AI — that is, AI-enabled robots capable of learning — under real-world industrial conditions"
(verified, BMW Group, 27 February 2026).
The programme's specific use cases were not publicly detailed in the source reviewed. BMW's formal evaluation sequence — theoretical assessment, laboratory evaluation, test deployment, pilot — was explicitly described as applying to this programme (verified, BMW Group).
What the evidence confirms: BMW is running a structured multi-stage evaluation of a second humanoid platform in a European facility. The programme is early-stage — initial announcement of planned pilot, not a report of sustained operation.
What the evidence does not confirm: Operating hours. Task performance. Commercial outcome. Timeline from current stage to production deployment.
Procurement lesson: A sophisticated buyer's structured evaluation sequence is the best available model for procurement governance. BMW's four-stage process from theoretical assessment to actual pilot is directly transferable to any buyer building their own humanoid vendor evaluation framework.
Evidence classification: VERIFIED — BMW Group customer disclosure
Deployment Evidence Summary Table
Deployment | Platform | Sector | Operating evidence | Outcome evidence | Evidence level |
BMW / F.02 (Spartanburg) | Figure 02 | Automotive | ~1,250 hrs, ~90,000 components, ~30,000 vehicles | Customer-reported production support | Customer-disclosed |
GXO / Digit (Flowery Branch) | Digit | Logistics | Multi-year RaaS | 100,000+ totes (company-claimed) | Commercial agreement confirmed |
Mercedes-Benz / Apollo (Berlin) | Apollo | Automotive | Active production-environment testing | Testing confirmed; ROI not disclosed | Verified (Reuters) |
BMW / AEON (Leipzig) | AEON | Automotive | Structured evaluation sequence initiated | Too early for outcome data | Verified (BMW) |
Schaeffler / Humanoid | HMND 01 | Manufacturing | Planned Dec 2026–Jun 2027 | Not yet deployed | Verified (Reuters — planned) |
Friction, Risk and Unresolved Issues
What are the biggest risks when buying a humanoid robot?
The six primary risks for industrial buyers in 2026 are: commercial immaturity (most platforms have limited real-world operating history), unclear TCO (enterprise pricing remains undisclosed across major vendors), integration obligations (buyers typically bear significant site preparation and engineering costs), intervention requirements (human supervision or teleoperation is embedded in current commercial deployments), vendor resilience (the market is capital-intensive and vendor continuity is not guaranteed), and contractual dependency (exit rights, data ownership, and software portability are unresolved). The largest risk is not that the robot fails — it is that the buyer has no contractual mechanism to respond when it does.
Risk 1 — Technical maturity is uneven and insufficiently evidenced.
The IFR characterised the humanoid market in June 2026 as one where reliability and efficiency still need to be proven, with most applications still trials or prototypes (verified, IFR Executive Roundtable). Interact Analysis identifies task reliability, efficiency, multitasking, embodied-AI maturity, endurance, and safety standards as continuing commercial barriers (verified, analyst research). Buyers should not assume that a vendor's production ramp implies operational maturity at customer sites.
Risk 2 — Human intervention is embedded in commercial deployments, not disclosed by default.
Agility Robotics CEO Peggy Johnson stated in a published interview that fully AI-controlled operation is not the approach used for its heavy commercial robot in current deployments. Mercedes-Benz explicitly described teleoperation and augmented-reality knowledge transfer as part of Apollo's current deployment methodology. 1X's NEO product documentation includes scheduled expert supervision for tasks the robot does not yet know (per 1X product documentation; company-claimed). Buyers who do not ask about intervention rates will not receive unsolicited disclosure.
Risk 3 — Integration burden is frequently underestimated and undercontracted.
BMW's Figure 02 programme involved integration through standardised interfaces in BMW's Smart Robotics ecosystem, with early involvement of production IT, logistics, and site teams. Integration is a cost centre — and in most vendor contracts, it is the buyer's cost centre. Site preparation, floor modifications, safety zoning, ERP/WMS connectivity, and network infrastructure are routinely excluded from robot pricing.
Risk 4 — Commercial opacity creates structural information asymmetry.
Enterprise pricing for Figure 03, Digit, Apollo, and Atlas is not publicly disclosed. No independently audited uptime benchmarks exist across vendors. No standardised cross-vendor task-success data is publicly available. Independently verified five-year TCO figures for any major deployment remain absent from the public record. Buyers are negotiating their first humanoid contracts without the market data that ordinarily informs capital procurement.
Risk 5 — Vendor resilience is not established at commercial scale.
Apptronik has raised nearly $1 billion in total capital. Figure reported more than $1 billion in Series C financing at a $39 billion post-money valuation in September 2025 (company-claimed / financial disclosure). Capital raised is not commercial viability. Buyers should ask for cash runway, production financing commitments, support organisation scale, spare-parts supply chain maturity, and escrow or continuity arrangements in the event of vendor restructuring or acquisition.
Risk 6 — Lock-in and exit risk are structurally embedded in humanoid procurement.
The robot ecosystem — hardware, AI models, fleet software, deployment-specific training data, support contracts — creates compounding dependencies. A buyer who does not contract explicitly for data ownership, software portability, support continuity, and termination rights at the outset may have very limited options when the commercial relationship changes.
Risk Assessment Matrix
Risk | Probability | Impact | Evidence available | Buyer response |
Integration underestimated | Medium-High | High | Partial (BMW case) | Scope integration in contract; budget 1.5–2× vendor estimate |
Vendor failure / restructuring | Unknown | High | Limited | Escrow provisions; exit clauses; data portability |
Unplanned downtime | Unknown | High | Weak | Contractual SLA; backup protocol; downtime cost in TCO |
Five-year TCO overrun | High probability of deviation | High | Very weak | Scenario model; staged commitment; milestone payments |
Lock-in / exit | Medium | High | Partial | Explicit data, software and termination terms in contract |
Intervention rate higher than expected | Medium-High | Medium-High | Partial | Measured intervention KPIs in pilot acceptance criteria |
Competitive Platform Comparison
How should companies compare humanoid robot vendors?
Procurement should compare humanoid robot vendors on eleven dimensions of commercial readiness — not on specifications alone. The dimensions are: industrial deployment evidence, named customer references, operating-hour evidence, task-level evidence, commercial availability, pricing transparency, integration maturity, support model, production capacity relative to committed demand, data and lock-in risk, and contract transparency. Specifications are company-claimed by default and tell a buyer almost nothing about the vendor's ability to deliver, integrate, support, and economically operate a robot at an industrial site.
Vendor specifications are company-claimed by default and tell a buyer almost nothing about a vendor's ability to deliver, integrate, support, and economically operate a robot at an industrial site. This comparison therefore evaluates commercial readiness across eleven dimensions — not technical specifications — to help buyers identify which vendors have provided sufficient evidence to progress to formal due diligence. Determining technical superiority requires independent engineering assessment, which falls outside this report's scope.
A note on evidence limits: the majority of dimensions in this table are rated on the basis of publicly disclosed information only. Company-claimed figures are marked accordingly. The absence of a public figure is itself a procurement data point — vendors with fully disclosed commercial terms, customer references, and support structures reduce buyer risk materially.
Buyer-Side Platform Screening Matrix
Dimension | Figure AI (F.03) | Agility Robotics (Digit) | Apptronik (Apollo) | Boston Dynamics / Hyundai (Atlas) |
Industrial deployment evidence | BMW (sustained, customer-disclosed) | GXO (commercial, vendor-claimed tote volume) | Mercedes-Benz (production testing, verified Reuters) | Planned 2028 deployment (verified, Reuters) |
Named customer references | BMW Group | GXO, Schaeffler, Toyota, Mercado Libre (company-claimed) | Mercedes-Benz | Hyundai internal |
Operating-hour evidence | ~1,250 hrs (F.02, customer-disclosed) | Not publicly disclosed | Not publicly disclosed | Not publicly disclosed |
Task-level evidence | Sheet-metal handling (customer-disclosed) | Tote movement (company-claimed) | Component transport + quality check (customer-disclosed) | Not publicly disclosed |
Commercial availability | Available (enterprise quote) | Available (RaaS + purchase) | Available (enterprise agreement) | Enterprise programme (2028 target) |
Pricing transparency | Not publicly disclosed | Not publicly disclosed | Not publicly disclosed | Not publicly disclosed |
Integration maturity | Smart Robotics ecosystem (customer-disclosed) | Agility Arc platform (company-claimed) | Digital Factory Campus integration (customer-disclosed) | Not publicly disclosed |
Support model | Not publicly disclosed | RaaS-embedded (company-claimed) | Not publicly disclosed | Not publicly disclosed |
Production capacity | One robot/hour (company-claimed) | 10,000 units/year (company-claimed) | Not publicly disclosed | Not publicly disclosed |
Data / lock-in risk | Not publicly disclosed | Agility Arc dependency (company-claimed) | Not publicly disclosed | Not publicly disclosed |
Contract transparency | Not publicly disclosed | Multi-year RaaS (structure confirmed) | Not publicly disclosed | Not publicly disclosed |
All vendor-originated data is COMPANY-CLAIMED unless otherwise noted. Verified or customer-disclosed classifications are noted where applicable. Absence of a public figure should be treated as a procurement question, not as confirmation of non-disclosure.
Strategic Recommendations — Automation and Procurement Buyers
Should a company buy, pilot or wait for a humanoid robot?
Apply the Green / Amber / Red decision gate. Green — the vendor provides named customer references, operating-hour evidence, measurable task performance, transparent TCO, defined support obligations, contractual acceptance criteria, and clear exit provisions. Proceed. Amber — the vendor has credible demonstrations and limited real-world evidence, but incomplete economics, significant human intervention, or unclear production-scale support. Structured pilot only. Red — the vendor relies primarily on demonstrations, future production promises, undisclosed pricing, or vague customer references. Do not commit capital.
A structured procurement process for humanoid robotics follows six stages. Execute them in sequence. Skipping any stage increases the probability of a failed deployment or a commercially disadvantageous contract.
Stage 1 — Define the use case before contacting vendors.
Identify the specific task or task cluster the robot will perform. Quantify current cost, throughput, quality rate, and labour structure for that task. Define measurable acceptance criteria — the performance the robot must achieve at the end of the pilot for the deployment to proceed. Do not allow vendor demonstrations to define the use case for you.
Stage 2 — Require evidence before issuing an RFQ.
Before formal procurement, ask each vendor for: named customer references willing to discuss operating performance; operating-hour data for the production version of the platform; task-level performance evidence in a comparable industrial environment; intervention rate data from current commercial deployments; and support infrastructure detail. This is not due diligence — it is the pre-screening that determines which vendors qualify for due diligence.
Stage 3 — Build the commercial model from the vendor quote, not from market data.
Request a complete five-year TCO breakdown from each vendor, including hardware or RaaS, integration, software, maintenance, spare parts, training, and support. Build three scenarios — downside, base, upside — using the framework in this report's Economics section. Do not use publicly available competitor pricing as a proxy, as enterprise pricing for the major platforms is not publicly disclosed.
Stage 4 — Structure the pilot with defined acceptance criteria.
A pilot without contractual acceptance criteria is a demonstration with a higher budget. The pilot agreement should specify: task performance KPIs; uptime and availability targets; intervention rate thresholds; data ownership during the pilot; what happens if acceptance criteria are not met; and the pathway from pilot success to production commitment.
Stage 5 — Contract the downside explicitly.
Before production commitment, the contract should address: SLA and uptime guarantee; replacement or repair response time; spare parts availability and lead times; software update and support obligations; data ownership and portability; termination rights and exit conditions; what happens if the vendor is acquired, restructures, or exits the market; and escrow or continuity arrangements for software and deployment-specific AI models.
Stage 6 — Apply the Green / Amber / Red decision gate.
Humanoid Vendor Decision Gate
Vendor identified ↓ Deployment evidence sufficient? NO → WAIT YES ↓ Economics viable in downside scenario? NO → WAIT YES ↓ Integration and support contractually defined? NO → PILOT (structured, milestone-gated) YES ↓ Contractual protections in place? NO → PILOT (negotiate contract first) YES ↓ GREEN — Proceed to production commitment |
GREEN — Commercially credible: Named customer references with operating evidence; measurable task performance; transparent TCO; defined support obligations; contractual acceptance criteria; clear exit provisions. Proceed to production commitment.
AMBER — Pilot only: Credible vendor with limited real-world operating history; incomplete economics; significant human intervention embedded; unclear production-scale support; insufficient independent performance evidence. Proceed to a structured, milestone-gated pilot — not production commitment.
RED — Do not commit: Vendor relies primarily on demonstrations; future production promises; undisclosed pricing; vague customer references; unsupported autonomy claims; no contractual performance guarantees; no clear failure or exit mechanism. Do not commit capital.
Buyers who apply the six-stage framework and the Green / Amber / Red gate to their next vendor conversation will know, by the end of that meeting, whether to issue an RFQ, propose a structured pilot, or walk away — and that is the decision this report was built to support.
Executive FAQ
1. What should I ask a humanoid robot vendor before signing a contract?
Ask for: named customer references willing to discuss actual operating performance; measured uptime, task success rates, and intervention rates from production deployments; a complete five-year TCO breakdown including all cost categories; defined SLA with response time and replacement commitments; explicit data ownership and software portability terms; and a clear termination and exit clause. Any vendor unwilling to provide these before contract signature is signalling a negotiating dynamic, not a product limitation.
2. What evidence should a humanoid robot vendor provide before a production pilot?
The vendor should provide: at least one named customer reference in a comparable industrial environment; operating-hour data for the production version of the platform; task-level performance data from real deployments (not controlled demonstrations); a disclosed intervention rate from current commercial operations; and a production-scale support model with defined response commitments. Evidence from the laboratory version of the platform is not equivalent to evidence from the production version — ask specifically about the version being proposed for your deployment.
3. How do you calculate the ROI and TCO of a humanoid robot?
Build the TCO from the actual vendor quote, covering hardware or RaaS fee, integration, site preparation, software, connectivity, training, maintenance, spare parts, downtime, human intervention, remote support, insurance, and end-of-life provisions. Run a base, downside, and upside scenario. Compare the full five-year cost against the quantified value of productive output in your specific use case. No independently verified five-year TCO benchmark exists in the public domain as of August 2026 — which means every buyer must build their own model.
4. How should companies compare humanoid robot vendors?
Compare vendors on eleven procurement dimensions, not on specifications: industrial deployment evidence, named customer references, operating-hour evidence, task-level evidence, commercial availability, pricing transparency, integration maturity, support model, production capacity relative to committed demand, data and lock-in risk, and contract transparency. Specifications are company-claimed by default. A vendor with lower specifications but stronger deployment evidence and better contract terms presents materially lower procurement risk.
5. Should a company buy, lease or use RaaS for a humanoid robot?
The correct structure depends on utilisation, capital risk appetite, obsolescence exposure, and the degree of operational flexibility required. RaaS reduces upfront capital outlay and transfers some technology-obsolescence risk to the vendor, but may create higher long-term cost and deeper contractual dependency. Purchase provides capital ownership and potentially lower long-term cost, but requires the buyer to manage maintenance, support, and end-of-life risk. Leasing sits between the two. In a market where platform specifications are changing rapidly and multi-year support commitments from early-stage vendors are unproven, buyers should model all three structures against their specific utilisation assumptions before committing.
6. When should a company wait instead of buying a humanoid robot?
Wait when: the vendor cannot provide named customer references with real operating evidence in a comparable environment; the economics do not close in the downside scenario; the support model is not contractually defined; pricing is entirely opaque; or the proposed deployment does not have defined acceptance criteria and a contractual remedy if they are not met. Waiting is not passive — it should be structured as a monitoring posture, with defined evidence thresholds that, when met by the vendor, trigger re-evaluation.
Scope and Disclaimer
This report is published by Physical AI Journal, operated by Sekason Research Limited, London, United Kingdom (Company Registration Number: 14339910).
This report provides procurement intelligence and independent market analysis for automation buyers, operations leaders, integrators, and investors evaluating humanoid robotics adoption. It does not constitute legal, financial, investment, engineering, or professional advice of any kind.
Physical AI Journal does not evaluate, certify, or endorse the safety, technical conformity, reliability, or fitness-for-purpose of any specific robotic platform or deployment. No sentence in this report is intended to state or imply a view on the engineering soundness, safety certification status, or technical conformity of any platform named herein. Readers must obtain independent engineering, safety, legal, and financial advice before making procurement, deployment, or investment decisions.
Company-claimed figures are labelled as such throughout this report and should not be relied upon as independently verified facts. Buyers should verify all vendor-originated figures directly with the relevant manufacturer or vendor.
The physical AI and humanoid robotics market is subject to rapid change. Platform specifications, pricing, commercial terms, and deployment status may change without notice. This report reflects the information available to the research team as of August 2026.
Full disclaimer: physicalaijournal.org/disclaimer
References and Strategic Sources
BMW Group — BMW Group to deploy humanoid robots in production in Germany for the first time — 27 Feb 2026 — https://www.press.bmwgroup.com/global/article/attachment/T0455864EN/644966 — VERIFIED — Customer disclosure (Tier 1)
Figure AI — F.02 Contributed to the Production of 30,000 Cars at BMW — 19 Nov 2025 — https://www.figure.ai/news/production-at-bmw — COMPANY-CLAIMED (Tier 2)
Figure AI — Ramping Figure 03 Production — 29 Apr 2026 — https://www.figure.ai/news/ramping-figure-03-production — COMPANY-CLAIMED (Tier 2)
Figure AI — F.03 Arrives at BMW — 30 Jun 2026 — https://www.figure.ai/ — COMPANY-CLAIMED / Customer confirmation (Tier 2)
Figure AI — Figure 03 product specifications — Current — https://www.figure.ai/ — COMPANY-CLAIMED (Tier 2)
Agility Robotics — Digit Moves Over 100,000 Totes — 20 Nov 2025 — https://agilityrobotics.com/ — COMPANY-CLAIMED (Tier 2)
GXO Logistics — GXO Conducting Industry-Leading Pilot of Human-Centric Robot — 6 Dec 2023 — https://investors.gxo.com/news-releases/news-release-details/gxo-signs-industry-first-multi-year-agreement-agility-robotics/ — VERIFIED — Customer disclosure (Tier 1)
GXO Logistics — Industry-First Multi-Year Agreement with Agility Robotics — 27 Jun 2024 — https://investors.gxo.com/news-releases/news-release-details/gxo-signs-industry-first-multi-year-agreement-agility-robotics/ — VERIFIED — Customer disclosure (Tier 1)
Mercedes-Benz — Digital Factory Campus — AI and humanoid robots — 18 Mar 2025 — https://group.mercedes-benz.com/unternehmen/produktion/produktionsnetzwerk/mbdfc-humanoide-roboter.html — VERIFIED — Customer disclosure (Tier 1)
Reuters — Mercedes-Benz takes stake in Apptronik, tests robots in factories — 18 Mar 2025 — https://www.reuters.com/ — VERIFIED (Tier 1)
Reuters — Apptronik raises $520m with Google/Mercedes backing — 11 Feb 2026 — https://www.reuters.com/technology/humanoid-startup-apptronik-raises-520-million-with-backing-google-mercedes-benz-2026-02-11/ — VERIFIED (Tier 1)
Apptronik — Apptronik closes over $935m Series A — 11 Feb 2026 — https://apptronik.com/ — COMPANY-CLAIMED / Financial disclosure (Tier 2)
Reuters — Apptronik launches training hub and Apollo 2 — 30 Jun 2026 — https://www.reuters.com/ — VERIFIED (Tier 1)
Reuters — Humanoid to deploy up to 2,000 robots at Schaeffler — 13 May 2026 — https://www.reuters.com/business/humanoid-deploy-up-2000-robots-schaeffler-plants-2026-05-13/ — VERIFIED / Company-attributed deployment target (Tier 1)
Reuters — Hyundai plans humanoid deployment at US factory from 2028 — 6 Jan 2026 — https://www.reuters.com/ — VERIFIED (Tier 1)
Hyundai Motor Group — Atlas industrial specifications / CES 2026 — 7 Jan 2026 — https://www.hyundai.com/ — COMPANY-CLAIMED (Tier 2)
Unitree — G1 product / pricing page — Current — https://www.unitree.com/mobile/g1/ — COMPANY-CLAIMED (Tier 2)
1X — NEO order / pricing page — Current — https://www.1x.tech/order — COMPANY-CLAIMED (Tier 2)
Reuters — Unitree IPO / company profile — 10 Aug 2026 — https://www.reuters.com/world/asia-pacific/what-is-unitree-why-are-chinas-humanoid-robot-makers-racing-list-2026-08-10/ — VERIFIED (Tier 1)
International Federation of Robotics (IFR) — Humanoid Robots: Vision and Reality — 14 Aug 2025 — https://ifr.org/ifr-press-releases/news/humanoid-robots-vision-and-reality-paper-published-by-ifr — VERIFIED (Tier 1)
International Federation of Robotics (IFR) — Executive Roundtable — humanoid reliability / efficiency — 24 Jun 2026 — https://ifr.org/downloads/press_docs/2026_06_24_IFR_Executive_Roundtable_market_presentation.pdf — VERIFIED (Tier 1)
Interact Analysis — Humanoid robot production surges tenfold in 2025 — Jun 2026 — https://interactanalysis.com/insight/humanoid-robot-production-surges/ — VERIFIED — Analyst research (Tier 2)
Interact Analysis — Humanoid Robots: Large opportunity but limited uptake — Mar/May 2025 — https://interactanalysis.com/ — VERIFIED — Analyst research (Tier 2)
This report is backed by authoritative research, independent verification, and structured analytical methodology.
© 2026 Sekason Research Limited · Physical AI Journal · physicalaijournal.org
Decision Intelligence for the Robotics Decade
Comments