Global Embodied Intelligence and Humanoid Robotics Industry Report

Part 1 Industry Story

In February 2026, BMW published an operating record from a humanoid robot pilot at its Spartanburg plant in the United States. During ten months in 2025, Figure 02 supported production of more than 30,000 BMW X3 vehicles. Its assignment was to remove and position sheet-metal parts for welding. BMW reported more than 90,000 components moved and approximately 1,250 operating hours. These are figures for participation in a production process, rather than a claim that the robot independently built complete vehicles. They are customer-reported results, with fleet size and independently audited economics undisclosed. [H01]

The release did not disclose intervention or interruption rates, or a full cost breakdown. Its record supports accumulated work on a specific production task; it does not establish general autonomy or an audited return on investment. [H01]

Our interpretation is that this bounded assignment makes the case useful. A production task gives developers a place to measure whether perception, movement and handling work together repeatedly. It also gives the customer a clear unit of useful output. A robot reaching a component once and a robot reliably completing an assigned operation across a production program are different evaluation milestones. The published record supports discussion of the latter within this particular pilot, while leaving several commercial questions open.

Another route into commercial operation appeared in logistics. On 27 June 2024, GXO and Agility Robotics announced a multiyear Robots-as-a-Service agreement for Digit at a SPANX fulfillment facility, following a late-2023 pilot. The described work involved moving totes from collaborative robots onto conveyors, coordinated through Agility Arc. The companies characterized the agreement as an industry first; this report treats that superlative as their claim. The announcement supplied neither the contract value nor a fleet count. [H02]

The analytical significance of the RaaS arrangement is that it ties a robot to an ongoing operating service. For a customer considering such an arrangement, the relevant comparison is the cost and reliability of completing the assigned work over time. Equipment, software and support need to be considered together. The agreement establishes a commercial structure; it does not, by itself, establish a favorable cost comparison with every other way of handling the same totes.

The BMW story also continued. On 25 June 2026, the automaker announced a Figure 03 project for logistics sequencing: picking components from containers, arranging them in a trolley, and handing onward transport to established automation. The release described the next application in forward-looking terms. It did not publish a completed operating record equivalent to the earlier Figure 02 pilot. [H03]

Taken together, these cases suggest the organizing question for this report: which physical tasks can an intelligent robot perform dependably, under what operating conditions, and at what total cost? Embodied intelligence concerns systems that connect perception and decision-making to physical action; a humanoid is one possible physical form. Keeping those scopes distinct allows us to examine the broader model and component ecosystem while testing humanoid business claims against specific customer applications. The industry's progress should be assessed through the relationship between technical capability and useful work, with demonstrations, pilots, commercial agreements and completed deployments identified separately.

Part 2 Industry History

The history has two connected strands: the development of machines with humanlike movement and the development of intelligence that acts through a physical system. The stages below are this report's analytical framework. They organize verified milestones without implying that every company followed the same path, or that one architecture inevitably replaced another. Their practical purpose is to explain why today's competition includes mechanics, control, data, models and integration at the same time.

Research humanoids established the mechanical problem

Waseda University completed WABOT-1 in 1973. In its institutional history, the university describes it as widely regarded as the first full-scale humanoid robot. The system could walk on two legs, grasp objects and conduct simple communication in Japanese. Waseda's WABOT-2, introduced in 1984, brought perception and coordinated movement to a specialized musical task. The qualification “full-scale” matters: it defines the historical claim rather than establishing an unrestricted first in all robotics. [H04]

Honda began its humanoid robotics research in 1986, presented P2 in 1996 and P3 in 1997, and announced ASIMO on 20 November 2000. Its original release emphasized a smaller body and smoother walking through i-WALK control. The launch specifications listed a height of 1.2 meters and a mass of 43 kilograms, with workstation and handheld-controller operation. These were engineering achievements in a particular generation of platform. [H05]

Embodied intelligence also developed through a different research route

In his 1991 paper, “Intelligence without representation,” Rodney Brooks argued for building complete systems incrementally and testing them through real sensing and action. His behavior-based approach connected parallel activity layers directly to the environment; the paper described mobile robots operating in laboratory and office settings. It is one influential research route into embodied intelligence, rather than proof that Brooks invented the entire field or its terminology. [H18]

The methodological link is evaluation of a complete acting system under real constraints. Contemporary foundation models have different architectures; this report therefore compares tested capabilities across physical forms.

The DARPA challenge put hardware and autonomy into demanding tasks

DARPA's July 2013 Atlas announcement introduced a humanoid platform for teams advancing from the Virtual Robotics Challenge. The platform combined hydraulic joints with sensing hardware; team-developed software and human operators guided its operation. The program connected simulated preparation to physical disaster-response tasks. [H06] The 2015 finals took place in Pomona, California, on 5–6 June. DARPA's post-event account identifies KAIST's DRC-Hubo as the top finisher and describes the winning teams' work. [H07]

In our reading, the challenge illustrated the importance of system integration. Teams using Atlas could pair a shared platform with different software, operator interfaces and strategies. The prerelease emphasized falls, battery management and partial autonomy under degraded communications. [H06][H07] Competition results should be assessed within their tasks and rules; commercial operation also requires evidence on reliability, maintenance, integration and economics.

In 2023, foundation-model research expanded the learning question

PaLM-E, published in March 2023, embedded images and state observations into a language model to support embodied reasoning. Its output remained text: decisions or plans that a separate low-level policy or planner translated into physical actions. [H08] RT-2, published in July, took a more direct action-generation route by encoding robot actions as output tokens when adapting vision-language models. Its authors explicitly cautioned that web-scale pretraining improved semantic and visual generalization without creating new physical motions beyond the robot training data's skill distribution. [H09]

Open X-Embodiment, first published in October 2023, investigated learning across robot platforms. The version read for this report, dated May 2025, describes data from 22 robotic embodiments contributed by 21 institutions. It studies positive transfer in manipulation and explains that robotics interaction data remains limited in breadth compared with vision and language datasets. These counts describe a research dataset, rather than shipments or customer deployments. [H10]

Our interpretation is that these projects changed how to formulate the learning problem. A developer could investigate whether knowledge learned elsewhere helped a robot interpret a situation, plan a task or reuse skills. The three examples address different links in that chain: grounded reasoning, action generation and cross-platform experience. Their differences matter commercially. Strong language understanding does not establish movement coverage; shared data does not guarantee transfer to every body; and a successful benchmark does not establish continuous operating performance. The model's role and the evidence for its particular capability must remain explicit.

Policy began addressing a complete industrial ecosystem

China's Ministry of Industry and Information Technology published humanoid robotics development guidance on 2 November 2023; the underlying document was dated 20 October. It set goals for 2025 and 2027 spanning models, motion control, components, production, application demonstrations and governance. The ministry's explanation also identified weaknesses in parts, operating systems, complete products and the industrial ecosystem. These are policy goals and diagnoses, rather than measured proof that the targets were achieved. [H19]

From 2024, commercial trials supplied a different kind of evidence

Boston Dynamics moved from hydraulic Atlas to an electric platform in April 2024. Its official introduction described a staged application-testing program beginning with Hyundai and emphasized the surrounding requirements of industrial deployment, including workflows and infrastructure. The current page does not display a precise publication day, so this report keeps the event at month-level resolution. [H11]

The GXO–Agility agreement described in Part 1 supplied a contractual example of RaaS. [H02] BMW's August 2024 Figure 02 trial supplied customer-side evidence from several weeks of sheet-metal handling tests. At that announcement date, BMW explicitly stated that no Figure robots remained at Spartanburg and that no return timetable had been established. [H17] The later 2025 pilot record belongs to a subsequent stage. [H01]

In 2025 to 2026 model ecosystems and product programs advanced together

DeepMind introduced two complementary models in March 2025: Gemini Robotics, a VLA with physical-action outputs, and Gemini Robotics-ER, which provides spatial reasoning and connects to existing low-level controllers. The announcement described primarily ALOHA 2 bi-arm training for the VLA, adaptation to other platforms and work with Apptronik. [H12] NVIDIA released the customizable GR00T N1 that month, with reasoning/action architecture and post-training for particular robots or tasks. [H13] In September, DeepMind introduced Robotics 1.5 and Robotics-ER 1.5 for action generation and reasoning, and high-level planning and tool use, respectively. At launch, ER had developer API access while the VLA was available to selected partners. [H14]

The hardware programs also developed. Boston Dynamics unveiled the product version of Atlas on 5 January 2026 and announced immediate manufacturing, with 2026 deployments scheduled for Hyundai's RMAC and Google DeepMind. Scheduled fleets are prospective commitments in that release. [H15] BMW's June Figure 03 project described a next logistics application. [H03] NVIDIA's March 2026 release announced GR00T N1.7 and previewed N2 for expected year-end availability; the preview alone cannot establish that N2 had launched by this report's cutoff. [H16]

Part 3 Market Landscape

Embodied intelligence describes a system that perceives, decides and acts through a physical body; a humanoid describes one possible body configuration. The distinction matters commercially. Google DeepMind tests transfer across ALOHA 2, a humanoid and a dual-arm platform, while Physical Intelligence studies mobile manipulation in unfamiliar homes. Neither line of work limits the intelligence layer to a two-legged machine. This report therefore treats robot models, development infrastructure and physical components as an enabling value chain, and distinguishes that broad ecosystem from narrower humanoid shipment statistics. [H14][V05]

The upstream, midstream and downstream divisions below are the report's analytical classification. Upstream supplies computation, sensing, motion hardware, models and training resources. Midstream combines them into a robot with a usable control system. Downstream turns that robot into a functioning workflow and maintains it. The industry mindmap follows industry → upstream/midstream/downstream → subsegment → representative company or product. Its lines indicate categories. They do not establish that a component manufacturer supplies every robot company shown elsewhere in the diagram. Several companies occupy more than one category because they develop both hardware and intelligence.

Models computation and data

At the intelligence layer, NVIDIA's Isaac GR00T combines customizable robot models with simulation and data-generation tools. Its N1.6 research release includes evaluation across multiple real robot platforms. Google DeepMind separately combines embodied reasoning with a vision-language-action model, while Physical Intelligence develops generalist manipulation policies. These are different technical and distribution approaches rather than interchangeable finished robot products. Model availability also requires precise language: access to an API, downloadable weights and participation in a restricted testing program confer different rights and capabilities. The report does not infer a commercial licensing fee or model-market revenue from a research release. [H13][V02][H14][V05]

Onboard computation is another product layer. NVIDIA announced general availability of Jetson AGX Thor and the T5000 module in August 2025. Its documented stack spans hardware interfaces, real-time control, perception and planning, and higher-level reasoning. This makes the platform relevant to both manufacturers and developers integrating models into machines. Purchasing a computer does not complete the sensing, motor control, power or safety system. The business role is supplying computing hardware and development infrastructure. Responsibilities for the resulting machine and its application still need to be defined. [V03]

Data is also becoming a product and service. AGIBOT's official DaaS menu lists customized data, customized collection environments, collection-system rental and thematic subscriptions. Its AGIBOT WORLD 2026 description discusses synchronized multimodal recordings, teleoperated collection and annotated recovery from errors. These materials demonstrate a data-service role alongside the company's robot role. They do not reveal that service's revenue, customer savings or profitability. Open data should be assessed under its specific license; a public announcement does not automatically authorize unrestricted redistribution. [V12][V13]

Motion and perception

The physical chain begins with motion. Motors, transmissions, bearings, encoders, controllers and structural parts have to work as one joint. Harmonic Drive represents precision strain-wave transmission: its mechanism uses a wave generator, flexspline and circular spline. maxon's HEJ family represents an integrated quasi-direct-drive approach using planetary gears and torque sensing. These examples illustrate alternative building blocks, rather than one universal joint architecture. A manufacturer can purchase components, buy an integrated actuator or develop its own assembly. The relevant customer value is controllable motion within a mass, thermal and durability budget, with performance maintained through repeated use. [V06][V07]

Perception has several distinct roles. Orbbec's Gemini 330 depth cameras and Jetson Thor integration represent visual sensing and its interface with computing. ATI's six-axis force/torque products measure forces and moments, while GelSight's DIGIT is a tactile research sensor for learning manipulation. These categories support different information: the scene, forces at a mechanical interface, and local contact. DIGIT the sensor should not be confused with Agility's Digit robot. Including these firms in the mindmap documents their product categories, not a verified supply contract with the humanoid manufacturers named in this report. [V09][V30][V08]

Robot manufacturers and applications

At the complete-machine layer, research equipment and industrial systems need to be separated. Unitree's G1 and G1 EDU specifications show different configurations and development access; a basic hardware offer cannot be treated as the price or capability of a fully equipped research platform. AGIBOT combines models with multiple physical configurations, and its June 2026 announcement reports a cumulative production milestone. UBTECH's industrial Walker S series focuses on handling, sorting and inspection, with battery management integrated into the system. The three examples show different routes into a developing market; their published product portfolios do not establish comparable end-user productivity. [V10][V11][M08]

Other manufacturers illustrate the importance of application engineering. Figure's BMW deployment involves a defined production task. Agility combines Digit and Arc with GXO's existing automation. Boston Dynamics' Atlas materials describe robot, fleet software and enterprise-system integration, alongside an early-customer deployment journey. Selling an industrial humanoid therefore extends beyond delivering a body: the supplier has to translate a customer's process into sensing, actions, exception handling and maintenance. Specific stage labels should be preserved—trial, contracted commercial deployment and a wider rollout are different claims. [V16][H02][V19]

The household route introduces another operating model. 1X's NEO launch describes basic initial autonomy supplemented by expert guidance for unfamiliar tasks. Tesla's second-quarter 2026 filing, meanwhile, describes continued Optimus development and preparations and investments for large-scale production. Both are participants at the integrated-machine layer, but these disclosures are different forms of evidence. A preorder offer demonstrates an intended customer proposition; production preparation demonstrates investment in future capability. Neither should be converted into an independently verified installed base or profitable autonomous labor service. [V20][V22]

Delivery and business models

Downstream is where these technologies acquire an economic purpose. Manufacturers need reliable loading, handling, sorting and inspection. Logistics operators need dependable transfer between equipment already in place. Research and education customers need development access and repeatable experiments. Household buyers need useful assistance in variable spaces. These are distinct customer jobs, rather than a single interchangeable market. An integrator, fleet operator or service team connects the robot to the actual work: defining the task, arranging interfaces, accepting the installation, managing exceptions and supporting operation. The value created must be demonstrated for each application.

Three business-model categories help organize the evidence. Equipment sales transfer a machine and associated obligations under a specified package. The GXO–Agility agreement illustrates Robots-as-a-Service; AGIBOT offers data services. These categories do not imply identical risk allocation: buyers still need to examine who pays for integration, downtime, remote assistance and replacement parts. An appealing subscription price can coexist with substantial implementation work, just as a low hardware price can coexist with an expensive development project. [H02][V13]

The report's analytical conclusion is that competition occurs at several levels simultaneously. Component firms compete on controllable performance and supply quality; infrastructure firms on development and deployment capability; robot makers on a usable machine; and application providers on an accepted outcome. A manufacturer's role, a robot's body shape and its operating maturity should therefore be shown separately. For downstream purchasing, the most informative comparison is the cost and quality of a completed job under the same conditions. A catalog specification, a cumulative production count and a successful demonstration each provide useful evidence, but answer different questions.

What the market boundaries allow the numbers to say

Embodied intelligence spans models, data, computing, physical machines and operating services. Humanoid robotics is one product category within that broader activity. The public materials actually read for this report do not establish a unified global denominator for company shares across the whole industry. We therefore use two separate windows into humanoid hardware: additional installations during 2025, and shipments during the first half of 2026. Each retains its publisher's definition and period. The charts cannot establish a company's share gain or loss between those windows, or represent the whole embodied-intelligence industry.[M01][M02]

Counterpoint estimates approximately 16,000 additional humanoid installations worldwide in 2025. Its original graphic assigns AGIBOT 31.9%, Unitree 26.5%, UBTECH 5.2%, Leju 4.9%, Tesla 4.7% and other vendors 26.8%. These measure annual additional installations, rather than cumulative installed stock or external sales revenue. The article discusses AGIBOT's X2 and G2, so wheeled products are present; a complete installation-counting methodology is not disclosed. The published one-decimal shares total 100%. The approximate overall volume must not be converted into exact vendor unit counts.[M02]

SAG's preliminary estimate for global humanoid shipments in H1 2026 is approximately 19,100 units. Its same-source graphic assigns AGIBOT 44%, Unitree 31%, Galbot 5%, UBTECH 4%, Leju 3% and others 13%, totaling 100%. Its definition covers full-size bipeds, smaller bipeds, wheeled robots and ultra-bionic robots; Leju includes KUAVO but excludes Aelos and Roban. All figures are rounded and preliminary. Displayed vendor volumes total 19,000, leaving a 100-unit difference from the published total. That does not authorize changing the published percentages.[M01]

A separately read Counterpoint article places H1 2026 shipments above 22,000. That differs from SAG's approximately 19,100, and the public articles do not disclose enough methodological detail to explain the entire difference. This report attributes each estimate separately, preserves its denominator and does not average or splice them. Early-market statistics are sensitive to product boundaries and counting methods. Year-over-year measures are interpretable only within a research provider's consistent series.[M03]

Customers and geography behind the rankings

Hardware rankings can conceal what a supplier actually sells and to whom. A research platform, a performance-rental robot, a data-collection machine and a robot moving materials in a factory may all appear in a provider's humanoid count. Their revenue structures, operating difficulty and customer acceptance requirements differ. Our analytical conclusion is that unit share helps describe manufacturing and channel scale; it cannot by itself measure industrial-automation capability or establish customer investment returns.

For the same H1 period, SAG estimates that Chinese vendors supplied more than 97% of global shipments and China accounted for more than 85% of demand. Vendor origin and demand location are distinct measures.[M01] A global assessment should therefore examine both China's manufacturing scale and how customers abroad procure robots. Demonstrating a product in many countries and generating recurring, verifiable overseas order revenue are different levels of evidence. A geographic label should describe the relevant measure rather than stand in for an entire competitive narrative.

Application categories also change the interpretation. Counterpoint's H1 2026 estimate leaves entertainment and performance, together with data production and research, above 60% of shipments. Intelligent manufacturing and warehousing and logistics account for approximately 13% and 5%, respectively. These describe Counterpoint's categories and cannot replace another provider's application assumptions.[M03] In our assessment, early research and performance orders can provide income, developer relationships and production utilization. Whether they become repeat industrial purchases must be examined through renewals, unattended operating time and the cost of supporting customer sites. A growing installed population can be commercially useful without yet proving generalized labor substitution.

Shipping, acceptance and profit are different stages

In January 2026, Unitree's official clarification described more than 5,500 pure humanoids in 2025 as actually sold and delivered to end customers, and reported production exceeding 6,500; it expressly excluded dual-arm wheeled products.[M05] Its later prospectus records production of 5,716, shipments of 5,511 and recognized sales of 5,215. It explains that some products had shipped under contracts but had not completed receipt, acceptance or other revenue-recognition procedures by year-end.[M07] The public materials do not fully reconcile the January production figure or the “sold and delivered” wording with the later table. Each disclosure should retain its own date and definition. Factory output, shipment, recognized sales and orders are distinct stages and cannot be added together to measure market size.

That prospectus reports RMB867.8319 million of humanoid main-business revenue in 2025, accounting for 51.78% of main-business revenue. The average humanoid selling price was RMB166,400, down 36.10%. This reflects the company's recognized sales and model mix; it is not a uniform global robot price.[M07] Company-wide revenue was RMB1,699.2693 million, consolidated gross margin 60.44%, and profit attributable to the parent after excluding nonrecurring items RMB590.7528 million.[M06] Quadrupeds and components also contribute to the business. Group profitability does not establish the same net margin for humanoids. The 51.78% denominator is main-business revenue and must not be applied directly to total company revenue.

UBTECH offers a different operating example. Its 2025 group revenue was RMB2,000.999 million. Full-size embodied-intelligent humanoid products and services generated RMB820.557 million, or 41.1%, with 1,079 relevant unit sales. Group gross margin was 37.7%, and group net loss RMB789.823 million.[M08] These figures do not calculate a loss per humanoid. Other robots and hardware sit within the group, the category includes services, and research and selling expenditure cannot simply be divided by those 1,079 units. Comparing companies requires matching products, accounting perimeter and the level at which profit is reported.

The latest readable interim announcement available by the research cutoff reports unaudited H1 2026 group revenue of RMB1,269.130 million, up 104.2%. Full-size embodied-humanoid products and services generated RMB590.299 million, or 46.5%, with 921 unit sales. Group gross margin rose to 44.7%, while group net loss remained RMB338.827 million. The disclosure also changed business categories and included newly acquired activities.[M09] Assessing improvement requires considering revenue, gross profit, operating expenditure and changes to the company perimeter together. H1 unit sales cannot be compared directly with the previous full year to calculate annual growth; a company's internal revenue proportion is not market share.

Funding scale and commercial revenue belong on separate lines

In September 2025, Figure announced more than USD1 billion of committed Series C capital at a USD39 billion post-money valuation, with manufacturing, training infrastructure and real-world data collection among the intended uses.[M10] In February 2026, Apptronik announced a USD520 million Series A-X extension. Together with the earlier USD415 million Series A, its Series A exceeded USD935 million, bringing total capital raised close to USD1 billion. Its announcement names Mercedes-Benz, GXO, Jabil and Google DeepMind as partners.[M11] These disclosures establish access to expansion resources. They do not establish operating revenue, customer spending or successful utilization of manufacturing capacity.

An accumulated round total must not be added again to an extension already included in that total. Likewise, valuation incorporates investors' expectations about future scale, intellectual property and execution; it does not promise a multiple of current profits. Our assessment is that competition unfolds simultaneously through manufacturing cost, model development and customer delivery. Capital can extend the period available for validation, but it cannot replace validation at operating sites. Regional comparisons should distinguish completed sales, contracts, pilots, capacity targets and financing rather than using one category as a proxy for another.

A useful commercial assessment consequently asks what the financing enables and what still needs to happen before it becomes revenue. Factory investment needs products that customers will accept. Training investment needs capabilities that work within customer constraints. Data collection needs rights, quality control and a path into improved performance. These are analytical questions, not claims that every funded project will succeed or fail. Applying the same questions to all companies makes the comparison more informative than ranking fundraising totals alone.

Customer sites are defining what can be charged for

The customer-confirmed GXO agreement described in P1 gives a concrete RaaS example. Published terms leave price, fleet size, service levels and payback undisclosed.[H02] The announcement establishes a contract and a defined workflow. Replicating it across facilities still depends on local conditions, integration work and sustained support. An operating agreement is stronger evidence than a demonstration, while remaining short of a published industry-wide return on investment.

BMW confirmed in June 2026 that Figure 02 supported the production of more than 30,000 BMW X3 vehicles over ten months in 2025, inserting sheet-metal parts within a welding process. A follow-on Figure 03 project moves to component sequencing in logistics, connecting with other transport automation.[H03] Supporting vehicle production does not mean independently manufacturing whole cars. The case is valuable because it identifies a work step, interfaces and a subsequent task; vehicle throughput must not become a substitute for robot sales. Customers buy a workable contribution to their process, rather than an abstract claim of generality.

Our assessment is that the commercial product may be an operating workflow: hardware, task software, maintenance, remote support and interfaces with customer systems. An equipment sale centers on delivery and acceptance; RaaS centers on continuing availability and operating cost. These structures allocate responsibilities differently. A low purchase price alone does not show that the customer receives a low-cost service, and a subscription alone does not establish that the supplier has attractive recurring margins. The relevant evidence includes who owns the equipment, who responds to faults and what constitutes accepted performance.

For a customer, choosing a humanoid, a wheeled dual-arm machine or specialized automation requires comparing safety, throughput, space, maintenance and total cost for the same task. Human-like appearance does not perform that comparison. Existing automation may remain part of the process even when a humanoid is introduced. Industry competition should therefore be judged increasingly by reliably delivered, billable work under stated conditions. This framework leaves room for different physical forms and business models without pretending that every promising prototype has reached the same commercial stage.

This report excludes Omdia's January 2026 version of its 2025 vendor-share graphic. The associated shipment table has an internal 50-unit discrepancy, and later first-party Unitree disclosures exceed its earlier estimate. The original report and graphic were actually checked; the exclusion is documented in Supporting Evidence.[M04][M05]

Adjacent markets and different robot definitions

Established automation provides an alternative against which humanoid systems must be assessed. IFR reports around five million industrial robots in operation in 2025 and more than 600,000 new installations that year. These are industrial-robot stock and flow, with no measurement of embodied AI adoption. [R01]

IFR separately reports about 7,000 full-size humanoids above 140 cm sold worldwide in 2025 for commercial and professional use beyond research and entertainment. That narrower boundary cannot be reconciled by subtraction with the broader humanoid estimates above. [R02]

Figure 3 describes professional service robot applications within IFR’s producer sample. Its purpose is to show established task markets where new systems face existing alternatives. The sample is not projected to the industry and changes annually; the chart neither counts humanoids alone nor classifies AI adoption. [R03]

Part 4: Industry Challenges

The difficult transition is from a successful action to a dependable service. For this report, commercial maturity is assessed through a chain of practical questions: can the robot perform the required job, recover when it fails, maintain that performance over a working schedule, operate safely in its actual surroundings, and produce an economic benefit after all associated costs? The analysis below combines documented capabilities and limitations with proposed evaluation criteria; it does not place every company at the same stage. Progress on one measure should not automatically be treated as progress on all the others.

Generalization and training data

Generalization remains an engineering boundary. Physical Intelligence's π0.5 paper reports manipulation in homes outside its training environments, but identifies persistent difficulties with unfamiliar handles, physical opening resistance, occlusion, distracted subtask selection and limited context. These are research results with stated limits. The practical implication is that a buyer should define which variation the system has actually been tested against. A useful evaluation changes objects, their placement, lighting, access and task order, then records completed jobs and assistance. This report does not equate an unfamiliar test environment with open-ended competence in every home. [V05]

Data volume needs to be judged by what it represents. NVIDIA's N1 announcement reports generating 780,000 synthetic trajectories in an internal experiment; that demonstrates a scaling tool rather than proof that every relevant physical situation has been modeled. AGIBOT's dataset description retains error-recovery trajectories, illustrating that failures and corrective actions can form part of the learning record. Deployment decisions should examine scenario coverage, calibration, rare failures and real-machine validation alongside hours or episodes. Simulation provides a way to expand testing and training; its value must ultimately be checked against the task on physical hardware. [H13][V12]

Reliability power and useful output

Mechanical reliability can constrain otherwise capable software. Figure identifies the forearm as its leading hardware failure point at BMW and discusses compact packaging, dexterity and thermal constraints. It also defines cycle time, placement accuracy and intervention count as operational KPIs. Its published accuracy and intervention goals must remain goals unless achieved results are separately disclosed. The implication for procurement is to request distributions over normal shifts, not just a best run: repeated failures, repair time and performance drift can matter more than a demonstration's peak speed. [V16]

Payload and endurance require conditions. Unitree explicitly notes that arm load varies with extension posture and that specifications vary with configuration and scenario. Atlas distinguishes an instantaneous 50 kg capacity from a sustained 30 kg capacity and publishes a four-hour battery figure. Those measures cannot be pooled into a single ranking of industrial productivity. A useful acceptance test specifies reach, object geometry, gripping method, movement, duration and required accuracy. It also separates powered hours from productive hours, because replenishment, maintenance and interventions interrupt output. [V10][V19]

Battery exchange addresses one interruption, with its own infrastructure. UBTECH advertises autonomous Walker S2 battery exchange within three minutes using a dual-battery system and charging stations. That is a product capability claim, not a published audit of uninterrupted round-the-clock availability. Similarly, more computing power has an energy cost: NVIDIA lists a 40–130 W range for the T5000 module, which is only one part of a robot's power budget. A customer's assessment should include stations, spare batteries, charging capacity and the energy consumed by motors and computation under the actual workload. [V15][V03]

Safety throughout the operating lifecycle

Safety has to match the application. ISO 10218-1:2025 addresses industrial robot design; Part 2 addresses applications and integration. Their public scope excludes publicly accessible service and household uses and does not cover specified mobility hazards associated with mobile platforms. ISO 13482 addresses defined personal-care robot categories and excludes industrial and medical-device robots, among others. These are scope boundaries, not a reason to say humanoids have no relevant standards. A robot's humanlike shape does not by itself identify the correct safety framework or establish conformity. [V23][V24][V25]

Mobility and dynamic balance add concerns beyond a stationary arm. A3's October 2026 standards update discusses bystanders, stopping distance, protective and emergency stops, energy control, and continued safe operation after deployment. It describes dynamically stable humanoids within the industrial-mobile-robot landscape. A site assessment should consider the complete machine, its payload, its route and nearby people together. A new hand, a heavier object, a changed floor layout or a software update may change the conditions on which an earlier assessment relied. Safety is consequently a lifecycle responsibility. [V26]

Semantic safety and a dependable protective layer also have different functions. DeepMind describes both high-level safety reasoning and lower-level robot safeguards. Agility and FORT's October 2026 announcement describes a planned Digit 5 architecture spanning a pendant, onboard communication and offboard interfaces. The latter remains a memorandum and forward-looking development disclosure. These sources support the need to distinguish reasoning about a safe action from an implemented protective function; they do not establish that a foundation model or an announced partnership provides universal certification for nearby human work. [H14][V27]

Remote assistance and governance

Remote assistance is a service and cost boundary. NEO's order page explicitly describes basic autonomy and scheduled Expert Mode for tasks it does not know, alongside remote-control options. Its ownership and subscription offers are different packages, not interchangeable estimates of total operating cost. The buyer needs a clear definition of autonomous completion, remote supervision and direct intervention. This report recommends measuring assistance minutes per accepted task and asking what happens when connectivity or a remote operator is unavailable. For a household or a confidential factory, the same arrangement also raises concrete questions about access to live views, data retention, permission and auditability. [V21]

Policy should be discussed with equivalent care. The European Commission's explanation of product-related high-risk AI sets cumulative conditions involving a covered product or safety component and a required third-party conformity assessment. It does not classify every robot as high-risk merely because it uses AI. A practical assessment starts by documenting intended use and responsibility for the robot, AI functions and installation together; obligations may differ across industrial, consumer and other applications. Exact legal requirements must be checked against the applicable jurisdiction and product. [V28]

Market access and product regulation

Market access now affects the global comparison. On 28 July 2026, the FCC added foreign-produced advanced robotic devices to its Covered List, subject to a DoW conditional-approval exception. The public notice prevents covered equipment from receiving FCC equipment authorization. The official FAQ says the production test is country-neutral; company nationality alone does not determine coverage. It also distinguishes new models from previously authorized ones: the July addition itself does not prohibit continued importing or selling of authorized models, or lawful consumer use. The report therefore does not describe an indiscriminate ban on existing foreign humanoids. [V33][V34]

Coverage depends on device characteristics and manufacturing status. The FCC's FAQ describes qualifying ground-mobile machines, including humanoids and quadrupeds, with weight, environmental sensing, connectivity and controlling-software conditions; it excludes fixed, stationary industrial arms. A supplier's body shape or headquarters is therefore insufficient for a compliance conclusion. Conditional exceptions also exist in practice: on 18 September, the FCC exempted ANSCER's AR250, AR650 and AR1250 platforms, subject to continued compliance with an onshoring plan and updated vetting. For purchasers, model authorization and applicable exception status become part of procurement evidence. [V34][V35]

The EU timetable also requires version control. In its 27 July 2026 announcement, the European Commission states that AI Omnibus entered into force that day and that rules for high-risk AI embedded in Annex I physical products apply from 2 August 2028; the Annex III timetable is 2 December 2027. These dates are attributed to the Commission's policy explanation. The report did not independently read the amended EUR-Lex legal text and does not treat every humanoid as high-risk. The cumulative classification conditions in the Commission's Article 6(1) guidance remain the starting point for product-specific assessment. [V36][V28]

Unit economics and purchasing evidence

The economic test is total cost per accepted output. The report proposes an accounting framework that includes equipment or service payments, integration, site preparation, training and data, batteries and energy, maintenance and spares, software, supervision and insurance where relevant. Costs should be assigned without double-counting items already included in a service package. Accepted output should use the same quality standard as the alternative process. This is an author-developed assessment framework, not a disclosed company cost model or a forecast of industry margins.

A simple decision calculation is annual verified benefit minus annual recurring robot-related cost; a payback calculation is meaningful only if that net benefit is positive and the initial investment is defined. Benefit may include avoided paid work, additional saleable output or reduced rejects, but the same gain should not be counted twice. Availability, useful utilization, task quality and assistance affect the result. The comparison should include the available alternatives—a fixed arm, a mobile manipulator, existing automation, a redesigned process or human work—under the same required workload. No numerical payback or universal labor-substitution ratio is inferred here.

A cumulative shipment, factory rollout or operating-hours announcement does not provide all of those inputs. The report consequently recommends stage-specific evidence. For an early system, request repeatable task and recovery trials. For a pilot, request representative shifts, exception logs and measured integration effort. For a commercial service, request accepted output, actual assistance, downtime, full cost and contractual remedies. These measures connect the industry's advances to an inspectable customer result while preserving what remains undisclosed.

Story and history sources

Market and company economics

Models components and operating constraints

Adjacent market statistics

Market access and regulation

Cover photograph C01

Unitree G1 humanoid robot and Go2 quadruped on an exhibition stage at Japan Mobility Show 2025, Tokyo Big Sight, October 31, 2025. The photograph illustrates robot hardware and does not establish industrial autonomy.

Photo by RuinDig / Yuki Uchida, CC BY 4.0, via Wikimedia Commons. Reproduced without pixel changes.

Original image: https://upload.wikimedia.org/wikipedia/commons/5/5a/Japan-Mobility-Show-2025-RuinDig_0542.jpg

Photographed 2025-10-31; uploaded 2025-11-28; read 2026-10-12. Original resolution 3840 × 2160 pixels. Version v2 uses the unmodified original photograph, proportionally scaled in the document. English alternative text: A silver and black Unitree G1 humanoid in side profile, with a Go2 quadruped to its right, on a lit exhibition stage in Tokyo.

The photograph documents exhibition hardware, not autonomous factory performance. Optional body photographs were omitted; data charts and the value-chain map are separately authored from the recorded evidence.