Global Embodied Intelligence Industry Development Analysis Report 2026

Key Findings

Embodied AI is moving from proof of concept into the inflection point of industrialization and commercialization, and has become a strategic high ground in global technological competition and new quality productive forces.

It should be noted that this report focuses on robot-form embodied intelligence, covering humanoid and non-humanoid generalized embodied applications such as quadruped robots, collaborative arms/robotic arms, wheeled/AMR, and special robots; it does not include autonomous mobile unmanned systems such as self-driving cars, drones, and unmanned boats in the analysis scope.

Global industry development trends:

1. The industry has entered a commercialization elimination race, with a gap remaining between scale and delivery

In 2026 the market enters the "first year of mass production" and the inflection point of scenario commercialization. Looking back at 2025, global humanoid robot shipments were about 25,000 units, with Chinese manufacturers accounting for about 52%; global humanoid/embodied financing exceeded USD 11 billion, and China's embodied intelligence financing reached RMB 44.4 billion. By the first half of 2026, global humanoid shipments had reached about 19,000 units, with China's share rising above 97%, and volume, price, and delivery all heating up together. However, "hot orders, cold delivery" remains to be solved: the ratio of announced intended orders to actual delivery is only about 10%–20%, and commercialization remains the key bottleneck for the industry.

2. The world shows a regional pattern of "multi-polar parallel progress, staggered complementarity"

North America focuses on the "brain", computing power, and capital; China focuses on manufacturing, supply chains, and scenarios; Europe focuses on industrial rigor and safety standards; Japan and South Korea focus on precision manufacturing and system integration. No single dominant force has yet formed. National policies have shifted comprehensively from "encouraging R&D" to three-dimensional competition based on "procurement pull + scenario opening + standard-setting power".

3. Industry value is converging toward three layers: "brain—body—scenario data", and the driving forces are switching in tandem

The technology architecture is moving from a discrete "perception—planning—control" approach toward an end-to-end paradigm represented by VLA/world models; industry driving forces are shifting from a single catalyst of venture capital to three-wheel coordination of "real scenario pull + policy supply push + manufacturing supply chain support"; a three-layer enterprise ecosystem of foundation layer, product layer, and application layer has taken shape, and the focus of competition has moved up from "demo prototypes" to "scale delivery and continuous iteration". This layered structure directly determines the staggered positioning of China and the United States along the value chain.

4. China and the United States are positioned at staggered points along the upper and lower ends of the value chain, converging at real delivery

The essential threshold of the US path lies in the "brain" (foundation models, computing power, and capital), while the essential threshold of the Chinese path lies in the "body" (supply chains, mass production, and cost). Each side's shortcoming is precisely the other's strength, forming a complementary upgrading path. Real commercial deployments show that the "brain and system definition" of leading humanoid robots is still dominated by US and European companies, while "body and mass production capability" increasingly leverage Chinese and East Asian precision manufacturing supply chains to reduce costs—Tesla Optimus is the most typical example; however, there are also cases such as Agility Digit that deliberately take a US domestic supply chain route, showing that "brain in the US, body in China" is not the only path. Open complementarity lowers the cost curves of both sides, but at the four levels of technology, supply chain, standards, and market, a strategy of "competitive cooperation and selective openness" is still required.

Trends and Outlook:

1. The China-US dual-core pattern is difficult to reverse in the short term, but the dimension of competition is moving from "staggered" to "opposed"

The United States is accelerating efforts to shore up shortcomings in manufacturing and supply chains (through domestic factory construction and allied contract manufacturing), while China is launching breakthroughs in upstream foundation models and original algorithms (through in-house development by major companies and open-source ecosystems). In the next 3–5 years, the two countries will gradually move from staggered competition in which "each guards one segment" to head-to-head competition across the entire chain, at which point cost, efficiency, and scenario density will determine the outcome.

2. Europe's "research + governance" faces commercialization pressure and may upgrade toward "rule export + high-end manufacturing"

As embodied intelligence enters the stage of scale deployment, basic research and governance discourse alone can hardly support an industrial closed loop. If Europe cannot form a commercial closed loop in complete machines or core components, its "research + rules" route may be marginalized; conversely, if it can rely on ISO standard-setting influence and high-end precision manufacturing to transform "rule advantages" into a business model of "standard licensing + high-end components", it can maintain long-term influence upstream in the value chain.

3. The commercialization threshold is approaching, and scale deployment is beginning to materialize

2026 is at the starting point of the commercialization threshold; humanoid and generalized embodied products are moving from demonstration to mass production, and scenarios such as warehousing and logistics and industrial inspection are entering scale deployment. Morgan Stanley has raised its China shipment forecast twice; early-mover cities will irreversibly lock in talent density, while the cost for latecomers to catch up rises steeply over time, and the window period is narrowing.

4. The industry will enter a period of differentiation and reshuffling within three to five years

As the technology maturity curve passes its peak of expectations, capital and orders will concentrate on entities with real delivery capability; complete-machine capacity lacking a differentiated core will face clearing out, and the ranking of cities will also be rearranged accordingly.

5. Soft infrastructure and compliance capabilities are becoming future thresholds

As embodied systems enter scenarios such as elderly care and public spaces, functional safety, cyber, and ethical certification have changed from "bonus points" to "entry permits"; physical soft infrastructure such as computing infrastructure, green energy supply, and public testing and validation platforms, as well as institutional soft infrastructure such as data openness, standard certification, and governance rules, will determine a city's short-term attractiveness to enterprises and long-term incubation capability.

1. Overview of the Development of the Embodied Intelligence Industry

1.1 Definition and Industry Boundaries

Embodied intelligence describes an intelligence paradigm, not a particular hardware form; in principle, the same embodied model can run on different bodies. Accordingly, the embodied intelligence industry referred to in this report is not limited to humanoid robots, but covers humanoid and non-humanoid generalized embodied applications, including quadruped robots, collaborative arms/robotic arms, wheeled/AMR, wheeled + robotic arms, special robots, and other physical forms; it does not include autonomous mobile unmanned systems such as self-driving cars, drones, and unmanned boats.

Figure 1-1 Definition and Subdivisions of Embodied Intelligence

Autonomous driving and drones are excluded not because they are insufficiently intelligent, but because the problem structure differs: the operating environment of autonomous driving is highly structured, with rules given by traffic regulations and infrastructure and a single task objective; embodied intelligence faces semi-structured or even unstructured environments, where tasks vary with the scenario and frequent physical contact with people and objects is required. This distinction determines that the two follow different paths in data, safety certification, and business models. If the two are not distinguished, it will cause obvious misjudgments about the industrial scale and industrial chain form of embodied intelligence, and will also mislead corresponding development recommendations.

1.2 Value and strategic significance

The industrial value of embodied intelligence is essentially driven by the demand side. Around 2025, it concentratedly entered the core of policies of various countries; the direct motivation lies in the structural change in global labor supply (aging and user shortage), rather than purely technological evolution. Its strategic value is concentrated in four aspects:

Reshaping the production paradigm: directly addressing demographic constraints, replacing repetitive and dangerous labor, breaking the manufacturing employment bottleneck, and improving total factor productivity;

Expanding capability boundaries: going deep into extreme or high-precision scenarios such as disaster rescue, deep sea and aerospace, and medical rehabilitation, breaking through the physical limits of human physical strength and operational precision;

Reconstructing the industrial ecosystem: giving rise to a new industrial chain of "body manufacturing-core components-RaaS (Robot as a Service) operation and maintenance," and reshaping employment and value distribution structures;

Empowering scientific discovery: as an autonomous executor of experimental and engineering tasks, greatly compressing scientific research iteration cycles and becoming an "accelerator" for basic exploration.

From a long-term perspective, embodied capabilities have become a key dimension in the science and technology game among major powers. Whichever country is the first to achieve "reliable and affordable" embodied intelligence will grasp the right to define next-generation productivity tools and ecological leadership.

1.3 Development status

1.3.1 Development history of the embodied intelligence industry

Taking humanoid and mobile manipulator robots as the main line, its development can be divided into five stages:

Sprouting and automatic mechanical device period (15th-19th centuries): mainly automata driven by springs and water power, belonging to mechanical imitation, without perception and decision-making capabilities.

Foundation of modern robotics (1920s-1960s): the concept of robots entered the industrial field, teach-and-playback robotic arms appeared, and control theory took shape.

Academic humanoid research (1970s-1990s): represented by the WAP/WL/WABOT series of Ichiro Kato's team at Waseda University (from 1968), systematic bipedal walking research was carried out; in 1973, WABOT-1 became the world's first full-scale humanoid robot; this stage was mainly focused on motion mechanisms, with intelligent capabilities initially sprouting.

Dynamic motion era (2000s-2022): Honda ASIMO, Boston Dynamics Atlas/PETMAN, and others promoted highly dynamic motion control, and perception and planning capabilities improved rapidly, but the degree of intelligence was limited, and task-level generalization capabilities had not yet been formed.

Commercialization explosion (2023 to present): the combination of large models and embodied intelligence; Google DeepMind released RT-2 in July 2023, marking the formal establishment of the VLA (vision-language-action) paradigm and its becoming the mainstream technical route, with capital and policy entering simultaneously.

Maturity differs significantly across forms: industrial/collaborative robotic arms and AMRs have already been commercialized at scale, and humanoid robots are transitioning from factory and warehouse pilots to small-batch deployment; general-purpose products for home scenarios are still quite far from large-scale usability. The pace of progress in the past three years has been significantly faster than in the previous thirty years.

Sorting through the above history reveals a clear pattern: each round of industrial acceleration is triggered by the "spillover of underlying capabilities" from the previous layer-control theory spillover brought motion and balance capabilities, giving rise to industrial robotic arms; machine learning spillover brought perception and recognition capabilities, giving rise to vision and navigation applications; the current spillover is general large models, which significantly lowered the threshold for task understanding and instruction following, driving VLA and task-level generalization. However, the three foundational engineering constraints of motion control, reliability, and cost have not been automatically resolved as a result, and still need to be tackled one by one through engineering and manufacturing. This also explains why there is still a significant gap between "impressive demos" and "scale delivery."

2025 is the first year for embodied intelligence to leap from prototypes to mass production, with leading enterprises completing verification from laboratory prototypes to small-batch production and delivery; 2026 is generally regarded as the takeoff year for large-scale commercialization, with major manufacturers upgrading from sample-based delivery to the stage of scalable commodities with stable production lines, commercial orders, and on-site deployment.

Shipments: In 2025, global humanoid robot shipments were about 25,000 units, of which Chinese manufacturers shipped about 13,000 units, accounting for about 52% (Morgan Stanley); it is expected that by 2030 and 2040, global shipments will reach about 700,000 units and 46.4 million units, respectively, while shipments from Chinese manufacturers will reach 400,000 units and 22.3 million units, respectively (Morgan Stanley). At present, leading enterprises have already taken the lead in ramping up, and Chinese companies occupy a dominant position in global supply.

Sales: In 2025, global humanoid robot sales were US$3 billion, and it is expected that by 2030 and 2040, the global market size will reach US$40.9 billion and US$1.99 trillion, respectively (Morgan Stanley), with significant long-term growth space. In 2025, China's humanoid robot sales were about US$600 million, and it is expected that by 2030 and 2040 they will reach US$14.9 billion and US$465.8 billion, respectively (Morgan Stanley).

Financing heat: In 2025, global financing in the embodied intelligence field was about 670 deals and US$11.2 billion (KPMG). China's embodied intelligence field had about 358 financing deals in 2025, totaling RMB 44.4 billion; in the first four months of 2026, the financing amount had already reached RMB 52.5 billion (IT Juzi, People's Daily Online), with heat entering an accelerating upward channel.

Enterprise ecosystem: There are more than 300 humanoid robot body companies worldwide, and more than 140 complete-machine companies in China; as of the end of April 2026, there were approximately 2,861 existing embodied intelligence companies in China (CCID, Qichacha), highlighting the ecosystem clustering effect.

Commercial orders: In 2025, orders from leading companies began to show scale (e.g., Unitree, AgiBot and others ranked among the top globally in order value and shipment volume), but overall the sector is still in a transition stage from "impressive demos" to "scale delivery"; order fulfillment will become the key to differentiation in the next stage.

Figure 1-4 Global and China humanoid robot shipments (2025-2050F)

Figure 1-5 Global and China humanoid robot sales revenue (2025-2050F)

Figure 1-6 Global embodied intelligence investment and financing (2025-2050F)

Figure 1-10 Global humanoid robot shipment share by segment (2025)

Figure 1-11 China humanoid robot shipment distribution by market segment (2025-2026)

The embodied intelligence industry chain can be divided into three layers: the foundation layer, the product layer, and the application layer. The competitive logic and sources of barriers differ across these layers:

Foundation layer: covers core components (lead screws, reducers, torque/tactile sensors, joint modules, dexterous hands), chips and computing power, development platforms, algorithms and models, operating systems, etc. The barriers at this layer come from process accumulation, mass-production yield, and algorithm/model iteration. Chinese companies have cost advantages in structural parts and some transmission components, but there is still foreign dependence in high-precision six-axis force/tactile sensors, precision transmission materials (such as harmonic flexsplines), and training computing power—domestic substitution has achieved broad breakthroughs in mass-produced models, but high-precision, high-dynamic, miniaturized, and high-end scenarios such as medical/aerospace have not yet been fully conquered.

Product layer: centered on complete-machine manufacturing, humanoid body integration, and software-hardware engineering coupling; the competitive focus is mass-production consistency, BOM cost, and delivery pace. This is currently the segment where Chinese companies have the most obvious advantage, but it is also the segment with the thinnest profits: gross margins for upstream core components are generally 40%–70%, while complete-machine manufacturers that purely purchase and integrate components have gross margins of only 10%–20%, and most remain persistently loss-making. Therefore, starting in 2025, leading complete-machine companies have successively moved upstream into self-development (e.g., Unitree self-develops actuators/motors/control, Kepler's core component self-development rate exceeds 80%) or extended downstream into operations, in order to compete for higher-value segments.

Application layer: includes system integration, scenario operations, and RaaS services, etc., directly determining on-site data feedback and product iteration speed. Companies that master high-frequency real-world scenario data have the strongest long-term bargaining power—once the data flywheel starts, it will form a positive cycle of "more deployments → more data → faster iteration → lower costs"; however, most scenarios are still in the pilot stage, a stable profit model has not yet formed, and pure integrators lacking data accumulation have limited bargaining power.

Profit distribution among the three layers is uneven. Companies that master key components (high-power-density actuators, high-precision force/tactile sensors) and high-frequency scenario data have the strongest bargaining power; the middle product layer is most easily squeezed thin by scale competition, which is why complete-machine companies generally extend upstream and downstream, while external manufacturers such as Inovance Technology and Leaderdrive position themselves in core segments through vertical integration. For local governments, the marginal returns from supporting complete-machine assembly are declining, while positioning in key components, opening real-world scenarios, and participating in standard definition are more valuable.

(2) Technology maturity shows a gradient distribution

Different carriers of embodied intelligence are at significantly different commercialization stages, overall presenting a gradient pattern of "specialized first, general-purpose lagging; structured scenarios running through first, unstructured scenarios breaking through later."

First tier: industrial/collaborative robotic arms, AMR—already commercially scaled (★★★★★)

After decades of iteration, industrial robotic arms are highly mature in fields such as automotive, 3C, and metal processing; motion control, force-position hybrid control, and path planning have become standard capabilities. According to Interact Analysis data, global collaborative robot shipments in 2025 were approximately 57,000 units, with revenue exceeding US$1.2 billion; China, leveraging supply chain and price advantages, accounts for more than half of global shipments. In AMR, according to CMR Industry Alliance statistics, the global mobile robot (AGV/AMR) sales scale in 2025 was approximately US$6.3 billion, with sales of approximately 265,000 units, up about 16% year on year; among them, Chinese companies' sales were approximately 182,000 units, accounting for about 68.7% globally. SLAM, multi-sensor fusion, and fleet scheduling have reached industrial-grade reliability. The core feature of this tier is "mature technology—controllable cost—calculable returns," and it is the only form in embodied intelligence that has formed a large-scale commercial closed loop.

Second tier: wheeled + robotic arm, quadruped, special robots—transitioning from pilots to small-batch (★★★)

Wheeled + robotic arm combines motion economy, endurance, and manipulation capability, and is rapidly being deployed in semi-structured scenarios such as factory loading/unloading, warehouse sorting, and commercial inspection; AgiBot A2-W, Jingling G2, and others have already performed precision loading/unloading and inspection tasks on intelligent manufacturing lines, marking that "wheeled body + collaborative operation" is moving from demos to paid deployment.

In quadruped robots, according to industry blue book statistics, global quadruped robot shipments in 2025 were approximately 82,000 units (China about 65,000 units, accounting for 80%), with a global market size of approximately RMB 7 billion; Unitree (dominant in consumer-grade, research and education) and DEEP Robotics (dominant in industry-grade, special applications) form a "duopoly," while European and American manufacturers (Boston Dynamics, ANYbotics, Ghost Robotics) focus on high-end and military applications, with high unit prices but limited shipments.

Special robots (power inspection, pipeline inspection, nuclear environment operations) have clear task boundaries and strongly constrained environments, making engineering risks controllable; central state-owned enterprises such as State Grid and CNPC have begun regular procurement, but the total market is limited and tends toward project-based models.

Third tier: humanoid robots—moving from factory pilots to small-batch deployment (★★)

2025 is widely regarded as the "first year of mass production" for humanoid robots. According to IDC, global shipments were about 18,000 units and sales about US$440 million, up roughly 508% year on year, with Chinese manufacturers dominant. By manufacturer, Agibot and Unitree each shipped on the order of 5,000 units, together accounting for more than half of the global share, while overseas manufacturers such as Figure AI, Agility Robotics and Tesla each shipped around 150 units in 2025, forming a China-US division of labor in which "China leads mass-production delivery and the US sets the technical benchmark."

In 2026 the industry enters a "small-batch delivery" phase. Institutions such as TrendForce, SAG and Bank of America generally forecast global annual shipments of 50,000–100,000 units, but this figure remains an institutional forecast/capacity estimate rather than realized data, and a considerable share of orders consists of validation procurement and on-site training, which is not equivalent to expansionary demand. On the application side, industrial scenarios come first: automobile final assembly, 3C materials, warehouse handling and other tasks have clear objectives, standardized environments and strong cost-reduction demands, making humanoid value easier to quantify. But the core bottleneck remains the "brain"—if single-step reliability is 99% and each step is assumed to be independent and identically distributed, the overall success rate of a 100-step continuous task is only about 37%, still some distance from being a "stable worker."

Fourth tier: general household scenarios—early validation (★)

The home is the ultimate scenario for embodied intelligence and also the hardest: unstructured environments, long-tail tasks, and complex safety and liability boundaries. Current products remain mainly robot vacuums, companion robots and research platforms, while household humanoid robots with dual-arm manipulation capability are mostly in pre-sale, internal testing or forward-looking exploration (such as 1X NEO). The industry generally judges that before 2028 household scenarios will focus on data collection and developer validation, and that consumer-grade penetration may begin around 2030 as costs fall to the ten-thousand-yuan/ten-thousand-dollar level and safety certification matures, but widespread adoption will still take longer.

(3) Global division of labor

The geopolitical division of labor in the global embodied intelligence industry has taken initial shape, with major economies forming differentiated positions in different links of the industrial chain based on their own endowments.

China: focused on the "body"—manufacturing, supply chain and shipment volume. Relying on the world's most complete manufacturing system, China dominates complete-machine manufacturing, mass production of core components and large-scale shipments. In 2025, China accounted for more than 84% of global humanoid robot shipments, and leading enterprises such as Unitree, Agibot and UBTECH have formed delivery capacity on the ten-thousand-unit scale. Industrial clusters such as the Yangtze River Delta (reducers, servo motors) and the Guangdong-Hong Kong-Macao Greater Bay Area (complete-machine integration, consumer electronics supply chain) give China significant cost and efficiency advantages in "building robots, building them cheaply and building them fast." However, in "brain" links such as foundation models and high-end computing power, China is still catching up.

English text is being prepared.

Europe: deeply engaged in "rules"—functional safety, standards and certification. Europe has established the world's strictest entry thresholds at the institutional level, including functional safety, machinery directives and CE certification. Core standards such as ISO/TS 15066 and IEC 61508 are led by European institutions, and third-party certification bodies such as TÜV and SGS are authoritative worldwide. This "rule-maker" status means that although Europe does not have an advantage in complete-machine shipments, it holds the referee's whistle on global market access.

Japan and South Korea: deeply engaged in "precision"—core components and system integration. Japan and South Korea have deep accumulations in precision manufacturing and remain globally leading in core components such as reducers (Harmonic Drive, Nabtesco) and servo motors (Yaskawa, Panasonic), while established manufacturers such as FANUC and Yaskawa have decades of engineering experience in motion control and system integration. Although China is accelerating domestic substitution, the process barriers of Japan and South Korea in the most high-end and precise components remain significant.

Summary: the essence of the four-pole pattern is a staggered distribution of "manufacturing capability, intelligence capability, rule-making capability and precision capability." China excels in scale manufacturing, the United States leads in foundation intelligence, Europe takes the lead in rule-making, and Japan and South Korea stand on precision craftsmanship. In the short term this pattern is difficult to overturn, but China's supply chain advantage is extending upstream to precision components, the US foundation-model advantage is being diluted by the open-source ecosystem, Europe's rule-making advantage is being offset by local standard-building in emerging markets, and the precision barriers of Japan and South Korea are gradually narrowing amid the wave of domestic substitution. Over the next 3–5 years, the competitive-cooperative relations among the four poles will determine the final direction of the global embodied intelligence industry.

(4) Structural gap in commercial delivery

Industry estimates show that of publicly disclosed intent/framework orders, only about 10%–20% ultimately convert into firm contracts and paid deliveries. Among delivered units, display-type scenarios such as exhibitions, reception and research and education account for a relatively high share, while the share undertaking continuous productive tasks remains low, and some equipment is actually used for data collection and model training. The industry as a whole is still in transition from "demonstration and validation" to "scale delivery," and order fulfillment is the key variable at present.

1.3.4 Industry pain points and solutions

1.3.4.1 Core pain points and constraints

The commercialization bottleneck of embodied intelligence is not a single-point technical problem, but the result of capabilities, costs, data, supply chains and institutions being interconnected and mutually constraining. The core constraints on current industry development can be summarized at three levels: the hard threshold that determines "whether it can be used and whether it is worth using," the foundational capabilities that determine "how fast iteration can be," and the institutional channel that determines "whether large-scale procurement is possible." The three must be overcome in sequence and cannot be tackled with equal effort at once.

Level 1: Hard thresholds — determining "whether it can be used and whether it is worth using"

This is the most prominent pain point at present, and also the decisive factor in whether the industry can be commercialized.

① Insufficient reliability and generalization

Success rates for long-horizon tasks face severe probabilistic decay: if single-step reliability is 99% and each step is independently distributed, the overall reliability of a 30-step task drops to about 74%, and that of a 100-step task to only about 37%. In real tasks, errors propagate and cascade, while systems generally lack effective error identification and self-recovery mechanisms. At the same time, long-tail situations in real scenarios are difficult to exhaustively enumerate, and the industry has not yet seen a single-parameter Scaling Law comparable to that of large language models; capability improvements depend more on the coupling of data quality × hardware maturity × simulation closed loop × engineering refinement, rather than simply stacking larger model parameters.

② Shortcomings in dexterous manipulation hardware

The service life of dexterous hands under continuous operating conditions is generally short, and the weight of the complete unit remains higher than that of a human hand (about 400 grams for a human hand); under most operating conditions, their overall performance has not yet stably surpassed that of standardized grippers. The essence is insufficient reliability at the hardware level — operating precision, durability and cost have not yet all met standards simultaneously, constraining the sustained execution of complex tasks.

③ Unit economics below standard

Manufacturing customers typically require a comprehensive cost of use below US$15/hour and an investment payback period shorter than 2 years. At present, only a few leading companies in structured scenarios such as automotive and 3C are approaching this level under ideal operating conditions; in general scenarios, a stable positive margin has not yet formed among hardware depreciation, operations and maintenance, and the benefits of replacing human labor. In other words, the product "can be used" but is "not cost-effective," and customers lack the incentive to procure at scale.

Level 2: Foundational capabilities — determining "how fast iteration can be"

The three capabilities above the hard thresholds determine how quickly the industry can approach commercial standards.

④ Scarcity of real interaction data

The scale of high-quality embodied data is far smaller than that of text and image data, and collection costs are high and annotation is difficult. Constrained by the "simulation-to-reality" domain gap (Domain Gap), data-driven model iteration has not yet formed a positive closed loop of "deployment — collection — iteration," and the data flywheel is not running smoothly.

⑤ Immature engineering and supply chains

Insufficient mass-production consistency, yield and reliability constrain cost reduction through scale. In terms of core components, mid-to-high-end actuators, precision transmission materials (such as harmonic flexsplines), high-dynamic force/tactile sensing, and on-device training and inference compute still rely on foreign sources — domestic substitution has achieved broad breakthroughs in mass-produced models, but high-precision, high-dynamic, miniaturized and high-end scenarios such as medical and aerospace have not yet been fully conquered.

⑥ Insufficient supply of interdisciplinary talent

Engineering talent with simultaneous capabilities in algorithms, hardware and systems cannot keep pace with the speed of industry expansion. This constraint is masked in the short term by capital and orders, but in the long term may become a tighter bottleneck than model capability.

Level 3: Institutional channels — determining "whether large-scale procurement is possible"

Even if technology and economics meet standards, the absence of institutional entry points will still block large-scale volume growth.

⑦ Absence of standards, certification and safety

Humanoid and mobile manipulation robots lack a unified safety certification and testing system, and large-scale procurement lacks institutional entry points. Especially in scenarios with high safety requirements such as industrial, medical and household settings, the absence of certification is directly equivalent to the absence of procurement.

⑧ Distortion in capital and commercial validation

Some valuations and intent orders contain inflated elements, and the structural gap of "many framework orders, few firm deliveries" persists: among publicly disclosed intent/framework orders, the proportion ultimately converted into firm contracts and paid deliveries is about 10%–20%. Product-market fit (PMF) has not yet been fully validated by real paid deployments, and there is a deviation between capital-market expectation management and real commercial progress.

1.3.4.2 Solutions

In response to the above constraints, the embodied intelligence industry has formed a relatively clear set of solution paths, which can be summarized as the following eight:

① Scenario traction: start with standard industrial scenarios such as automotive, 3C, warehousing and logistics, then spill over layer by layer into general industry and commercial services;

② Architecture upgrade: use VLA / world models as the mainstream "brain" architecture to improve generalization capability and long-horizon robustness;

③ Hardware iteration and transition strategy: rapid iteration of dexterous hands and bodies, with smooth transition via mature forms such as standardized grippers and wheeled bases + robotic arms;

④ Data closed loop: real deployment collection + teleoperation teaching + synthetic data/simulation, to accelerate the data flywheel;

⑤ Cost reduction through mass production and domestic substitution: lower the comprehensive cost of use through large-scale mass production and supply-chain localization;

⑥ Lowering the bar for business models: use RaaS (Robotics as a Service) to convert customers' one-time capital expenditure (CAPEX) into operating expenditure (OPEX), lowering the adoption threshold for customers and feeding data back;

⑦ Standards and certification in advance: establish safety, testing and certification systems to open up channels for large-scale procurement;

⑧ Talent supply: joint university-enterprise training, with practical training bases and reskilling programs running in parallel.

The implementation of the above solutions requires systematic advancement within a collaborative framework of "enterprises — capital — policy — standards organizations — regional clusters": enterprises lead technology and product iteration, capital provides patient investment, policy fills institutional gaps such as first-purchase, standards, talent and cluster coordination, standards organizations and third-party evaluation ensure safety and mutual recognition, and regional clusters undertake supply chains and scenario implementation. The correspondence between each constraint and solution path, and the main driving entities, are shown below:

Table 1-1 Constraints and solutions for the development of the embodied intelligence industry

Overall, the eight solution categories are not advanced in isolation, but can be summarized into four interrelated main lines:

The technology main line (architecture upgrade, hardware iteration, data closed loop) addresses the capability ceiling problem of "making robots smarter and more reliable";

The commercial main line (cost reduction through mass production, lowering the bar for business models, scenario traction) addresses the commercialization speed problem of "making customers able to afford and willing to use";

The institutional main line (standards and certification in advance) addresses the volume ceiling problem of "giving procurement a basis and an entry point";

The element main line (talent supply) runs through the entire process and determines the sustainability of long-term industry iteration.

The four main lines have a sequential dependency in time: the technology main line determines "whether it can be done," the commercial main line determines "whether it is worth doing," the institutional main line determines "whether it can be sold," and the element main line runs through the entire process. At the current stage, the urgency of the commercial main line is higher than that of the technology main line—the technology gap can be closed through sustained investment, but if it cannot be proven that the products are "affordable and willing to be used," the industry will remain stuck at the stage of intention orders for a long time; the institutional main line, meanwhile, needs to be laid out in advance to avoid the situation where, once the technology and business mature, certification and standards block the channel for scaling up.

1.3.5 Development Trends of the Embodied Intelligence Industry

Technology evolution: from discrete modular control to a hybrid "VLA + world model" architecture

The technology foundation of embodied intelligence is undergoing a generational shift: the discrete modular architecture of "perception—planning—control" has fully shifted to an end-to-end paradigm represented by vision-language-action (VLA) large models; world models and "simulation-to-reality" (sim-to-real) technologies continue to narrow the gap between the virtual and the real, and collaboration between cloud training and edge inference has become the norm.

At present, VLA remains the mainstream technology route in the industry, but it has already shown multiple parallel evolutionary paths; Flow Matching has become the mainstream action generation paradigm, and the deep integration of reinforcement learning and VLA is driving models to leap from imitation learning to autonomous learning. However, VLA is essentially a "conditioned-reflex" agent—it learns the direct mapping of "do what you see" rather than the causal reasoning of "what will happen to the world after doing this"—and it is increasingly exposing structural bottlenecks in dynamic environment adaptation and long-horizon planning tasks.

As a result, technology evolution is moving toward the next stage: the World-Action Model (WAM). World models, by building interactive physical environment simulators, enable agents to predict future states and evaluate the consequences of actions, moving from "conditioned reflex" to "causal reasoning." It should be emphasized that this shift is a medium- to long-term evolutionary direction—in the short term, the industry will still rely mainly on VLA as the mainstream engineering architecture, and in the medium to long term it will evolve toward a hybrid architecture of "VLA + world model/WAM"; WAM represents the direction from mapping intelligence to prediction-reasoning intelligence, but it is still in an early stage. In the third-generation brain-inspired VLA architecture, deep integration of fast and slow systems has become a core design concept: the slow system is responsible for abstract reasoning and long-term planning (about 1–5 Hz), and the fast system is responsible for real-time perception and rapid response (20–100 Hz). Some frontier models show that embodied models improve nonlinearly after their parameter scale crosses the threshold of several billion, and whether the Scaling Law holds in the embodied domain still awaits validation with real-machine data.

In the longer term, technology evolution will continue to deepen along three main lines: from a single modality to a closed-loop mechanism of multimodal perception and interaction, from static scene understanding to dynamic environment prediction and adaptation, and from a single agent to multi-agent collaboration.

This shift means that knowledge production (papers, models, patents) and the ability to engineer research results into practical applications are increasingly becoming the fundamental source of the gap between cities—whoever can form a sustained iterative capability in foundation models and data closed loops will be able to take the initiative in the next stage of technological competition.

Industry progress and driving forces: from a single driver of capital to synergy among the three elements of "scenario-policy-manufacturing"

The industry is advancing rapidly along the three stages of "proof of concept—prototype demonstration—initial commercialization." 2025 is a year of buildup, and 2026 is a year of leap from technology accumulation to large-scale deployment. Morgan Stanley has twice raised its shipment forecast for Chinese humanoid robots (50,000 units in 2026, and 446,000 units/US$15 billion in 2030), indicating that China is moving from "prototype competition" into "early commercialization."

More critically, the structure of driving forces is switching from the single catalyst of early-stage venture capital to the synergistic superposition of three elements: real scenario pull, policy supply push, and manufacturing supply chain support.

Ecosystem structure: "chain leader" enterprises are deeply reshaping the supply chain, with open source and vertical integration proceeding in parallel

Embodied intelligence has formed a clear "three-layer ecosystem": the foundation layer (core components, actuators, chips, foundation models), the product layer (complete machine integration and humanoid bodies), and the application layer (system integration and scenario operations). Among them, "chain leader" enterprises are reshaping the supply chain with unprecedented depth. On the one hand, leading complete-machine enterprises, through strategic investment and joint R&D, are deeply binding key component suppliers such as core joints and dexterous hands, upgrading traditional procurement relationships into a "community of shared destiny" that shares risks and benefits; on the other hand, leading enterprises are accelerating vertical integration, bringing key links such as motors, batteries, semiconductors, and even sensors into their in-house R&D systems to achieve the greatest possible autonomy and controllability.

Ecosystem health no longer depends on single-point breakthroughs, but on the pulling power of chain leader enterprises, the density of supporting enterprises, and the degree of open collaboration. The open-source ecosystem is prospering in parallel—leading computing power platforms have open-sourced foundation models and training data, significantly lowering the innovation threshold for small and medium-sized enterprises and becoming an important engineering paradigm for the deployment of the embodied intelligence industry.

Scenario depth: from demonstration shows to "affordable large-scale deployment," with industry ahead of the home

Non-humanoid embodied forms (quadrupeds, collaborative arms, AMRs) have entered profitable large-scale applications, forming mature industry models in logistics, industrial, and other scenarios.

Humanoid complete machines, by contrast, are still in the triple validation period of "cost—reliability—ROI" and are moving from demonstrations into pilot commercial use in manufacturing, warehousing, services, and other areas. The real constraint has now shifted from "whether it can be done" to "whether it can be deployed at scale at an affordable cost," and cost competitiveness has become the core variable. In 2025, global humanoid robot shipments were about 18,000 units, with sales of about US$440 million (IDC, January 2026). The application path shows a clear logic of gradual penetration: automobiles and 3C electronics → warehousing and logistics → commercial services → home scenarios, while large-scale penetration of home scenarios will still have to wait until after 2030.

Business model: RaaS + large-scale manufacturing, with OPEX conversion and capacity ramp-up becoming the main battlefield

The business model is shifting from "selling equipment" to a dual-engine drive of "RaaS operations + large-scale manufacturing": RaaS converts customers' one-time capital expenditure into operating expenditure paid by the hour or by task, lowering the adoption threshold and continuously feeding back real operating-condition data; financing leases, insurance, government first-purchase programs and demonstration subsidies together underpin demand. On the supply side, leading global companies plan to break through 100,000 units of production capacity in 2026, and the focus of competition is shifting from project-based bidding and single-point demonstrations to mass-production consistency, yield rates and delivery cadence—"being able to build it" gives way to "being able to build it cheaply, stably and at scale." Embodied intelligence companies are evolving from hardware vendors into composite platforms of "model + hardware + operations," and the decisive factor has become safety, auditability and quantifiable ROI in real-world scenarios.

Regional and urban momentum: embodied intelligence becomes the core growth pole of "new quality productive forces"

At the regional and urban levels, embodied intelligence is widely regarded as a key track for future industries and "new quality productive forces," and urban competition has escalated from "competing for projects" to "a war to build ecosystems." Whoever can form systemic advantages in talent, capital, scenarios, supply chains and institutions will seize the new growth pole.

Globally, the number of embodied intelligence industrial parks and innovation clusters is increasing rapidly, and the industrial cluster effect continues to strengthen, making them an important vehicle for the transformation of regional economic momentum.

Spatial layout: deepening global clustering and a multipolar landscape

Embodied intelligence is highly clustered in spatial terms and multipolar: the United States has a dual core of the San Francisco Bay Area and Boston; Japan has formed a Tokyo-Nagoya-Osaka corridor; China presents a collaborative network of Beijing-Shanghai-Shenzhen-Hangzhou; and distinctive clusters in Seoul, Stuttgart and Zurich are rising rapidly. The formation of clusters depends on infrastructure such as computing power, green electricity, public testing platforms and open datasets.

The global division of labor presents a multipolar landscape of "US (brains)—China (bodies)—Europe (regulation)—Japan and South Korea (precision)": the United States leads in foundation models and computing power, China focuses on complete-machine manufacturing and large-scale shipments, Europe leads in functional safety and standards certification, and Japan and South Korea deepen their work in precision components and system integration.

1.4 Comparison of global regional embodied intelligence industry development

Global competition is not a race of fast and slow on the same track, but a contest in which different dimensions each excel: the United States is strong in foundation models, computing power and capital; China emphasizes manufacturing, supply chains and scenario implementation; Europe pays more attention to industrial rigor and safety standards; and Japan and South Korea focus more on precision manufacturing and system integration. The consensus among all parties lies in the common drivers of "labor shortage + AI maturity + capital influx," while the divergence is reflected in the United States competing in general intelligence and high dexterity, China breaking through in mass production and low cost, and Europe anchoring itself in safety and compliance. Policies of various countries have shifted comprehensively from "encouraging R&D" to three-dimensional competition of "procurement pull + scenario opening + standard-setting power."

Table 1-2 Development status of the embodied intelligence industry in various countries and regions

1.4.1 United States: a highland of technological originality and capital

Development status: Domestic companies such as Figure, Tesla and Apptronik, together with 1X (Norwegian background, US implementation), are clustered here, and the sector is in a stage of "technological leadership, mass production beginning."

Model and path: With "original research + venture capital + general AI capabilities" as the core, and relying on top universities and technology giants, it leads in embodied large models, open-source ecosystems (Open X-Embodiment, LeRobot, GR00T, π0) and capital density, with a path leaning toward "technology-driven, bottom-up."

Drivers and strengths/weaknesses: Breakthroughs in AI foundation models, GPU/cloud computing power, venture capital density and vertical integration by technology platform companies are the main drivers. Its strengths lie in VLA/world models, computing power and financing scale; its weaknesses lie in engineering costs, supply chain completeness and mass production cadence.

1.4.2 China: engineering and manufacturing scale advantages

Development status: The 2025 Government Work Report mentioned "embodied intelligence" for the first time; the outline of the 15th Five-Year Plan lists quantum technology, biomanufacturing, hydrogen and nuclear fusion energy, brain-computer interfaces, embodied intelligence and 6G as future industries. Beijing has proposed goals of 10,000-unit mass production and a 100-billion-yuan cluster, and Shanghai has established a 1-billion-yuan embodied intelligence industry fund. Companies such as UBTECH, AgiBot, Unitree and Yunji are advancing commercial pilots, among which UBTECH disclosed orders exceeding 1.3 billion yuan and deliveries of more than 500 units, AgiBot's 5,000th complete machine rolled off the line, Unitree pushed the price of consumer-grade humanoids down to 29,900 yuan, and Yunji has formed mature operations in hotel scenarios.

Model and path: Characterized by "manufacturing scale + scenario pull + policy aggregation," and relying on a complete supply chain and rapid engineering capabilities, it holds advantages in mass production cadence and cost control, with a path leaning toward "application-driven, top-down," while basic research is being accelerated to catch up.

Drivers and strengths/weaknesses: A complete supply chain, special policies in multiple provinces and municipalities (such as Shanghai's 1-billion-yuan fund) and the opening of public scenarios are the core drivers; for example, State Grid plans to centrally procure about 8,500 units of various embodied intelligence devices in 2026, with a total investment of about 6.8 billion yuan, covering power inspection, live-line work, emergency rescue and warehousing and logistics scenarios. China is strong in supply chains, mass production, cost and speed; it is weak in high-end dexterous hands and general-purpose "brains."

1.4.3 Europe: industrial robots and a rigorous approach

Development status: Germany (Franka Robotics) has a manufacturing base, Switzerland (ABB) has relevant hardware production capacity, and the United Kingdom has deep academic accumulation, forming the basic foundation. BMW has introduced Figure, Schaeffler has deployed NEURA, and Mercedes has piloted Apptronik Apollo; overall, the region is in a stage of "prudent deployment, high thresholds."

Model and path: Characterized by "industrial robots and research institutions + safety/standards first," with the Fraunhofer Society, the German Aerospace Center (DLR) and ETH Zurich leading frontier research, emphasizing reliability and compliance, and with a path leaning toward "steady and rigorous, standards first."

Drivers and strengths/weaknesses: Accumulated precision manufacturing expertise, ISO standards leadership, and the compliance framework established by the AI Act (with high-risk application provisions deferred to 2028), forming the "Brussels Effect." Europe is strong in precision manufacturing, safety and standards; it is weak in explosive growth in consumer-grade scenarios and in the iteration speed of new humanoid entrants.

1.4.4 Japan and South Korea: Robotics Traditions and Embodied Intelligence Redeployment

Development status: Toyota, Honda and Sony of Japan have accumulated experience over a long period, but humanoid commercialization has been slow; they are leveraging the rigid demand created by aging populations to advance care/rehabilitation robots. South Korea's Hyundai (Boston Dynamics) and Samsung are entering through capital and manufacturing capabilities.

Model and path: Characterized by "traditional accumulation + capital redeployment," with enterprises as the main actors and national strategy as the backstop; the path leans toward "gradual, enterprise-led."

Drivers and strengths/weaknesses: In 2026 Japan elevated "physical AI" to a national strategy; the Ministry of Economy, Trade and Industry's "AI Robot Strategy" set an implementation target of introducing about 10 million AI robots by 2040, with 387.3 billion yen allocated for fiscal 2026 and support of about 1 trillion yen over the next five years. In 2026 South Korea listed Physical AI as a national flagship program, targeting the deployment of 1 million intelligent robots by 2030 and the training of more than 10,000 specialists. Japan and South Korea are strong in component quality, precision manufacturing and system integration, and weak in general software ecosystems and cross-embodiment foundation models.

1.4.5 Other Emerging Regions

Southeast Asia and the Middle East have begun to lay out plans by leveraging manufacturing relocation and sovereign capital, but industrial chain supporting facilities, talent reserves and data closed loops remain incomplete. In the short term they will find it difficult to replicate the full-factor ecosystems of China, the United States, Europe, Japan and South Korea, and will play more the role of "manufacturing recipients + scenario testing grounds."

1.5 Differences Between the China and U.S. Paths and Their Respective Demonstrative Significance

1.5.1 Comparison of Path Difference Dimensions

In the new round of embodied intelligence industrialization driven by large models, China and the United States entered at almost the same time but took two completely different paths. The United States, relying on top universities, technology giants and dense venture capital, has formed a bottom-up model with "original technology + capital-driven + general AI" at its core; China, relying on a complete supply chain, a super-large market and strong policy coordination, has taken a top-down model characterized by "manufacturing scale + scenario traction + policy aggregation." The two paths show systematic differences in dimensions such as competitive focus, innovation logic and priority objectives, as follows:

Table 1-3 Comparison of China-U.S. Embodied Intelligence Path Characteristics

The differences in the above path characteristics are rooted in the deep divergence in the two countries' resource endowments and capability structures. The United States has advantages in originality in basic research, capital density and computing power platforms, but has structural shortcomings in hardware supply chains and mass-production processes; China leads in supply chain completeness, manufacturing scale and cost engineering, but still needs to strengthen its basic research accumulation and independent computing power, as follows:

Table 1-4 Differences in China-U.S. Embodied Intelligence Resource Endowments

In short, the essential threshold of the U.S. path is the "brain"—foundation models, computing power and capital density; the essential threshold of the Chinese path is the "body"—supply chains, mass production and cost engineering. Each is the other's weakness, and each is the other's direction for upgrading.

1.5.2 Characteristics of the U.S. Path and Its Demonstrative Significance

① For global innovation: Open foundation models and data flywheels significantly lower the threshold for global developers (open source such as GR00T, π0 and UnifoLM-VLA), alleviating the "embodied data scarcity" bottleneck; the first "robotic dexterity Scaling Law" has emerged (first-person data increased from 1,000 hours to 20,000 hours, and task completion rates doubled), proving that data scale remains the core variable for capability improvement.

② For vertical industries: Real paid validation of product-market fit (Figure achieved paid commercial use at BMW, at about $25 per robot-hour; Tesla reuses the FSD software stack to advance Optimus mass production) proves that "brain" capability is the key variable that creates the gap, pushing the industry from "video demonstrations" to "procurement decisions" and accelerating the formation of the technology-industry closed loop.

④ For ecosystem mechanisms: The RaaS model turns robots into nodes that continuously generate training data, forming a positive cycle of "deployment—data—iteration," which is the key mechanism for the ecosystem to mature.

1.5.3 Characteristics of the Chinese Path and Its Demonstrative Significance

China is characterized by "manufacturing scale + scenario traction + policy aggregation," relying on a complete supply chain and rapid engineering capabilities, and has advantages in mass production and cost. Its demonstrative significance is as follows:

① For global innovation: It proves that "a mass-producible body at a lower cost" is the other half of scaling—Unitree has pushed the price of an entry-level humanoid complete machine down to 29,900 yuan, shipped more than 5,500 units in 2025 (company disclosure, covering pure humanoids and excluding wheeled dual-arm models), and expanded overseas through channels such as Amazon, breaking the expectation that "humanoids must be expensive."

② For vertical industries: Mass production and cost reduction make automation affordable for small and medium-sized manufacturers, moving robots from being exclusive to big companies toward broad accessibility; State Grid plans to centrally procure about 8,500 embodied intelligence devices in 2026, covering four scenarios—power inspection, live-line work, emergency rescue, and warehousing and logistics—providing replicable automation templates for various industries.

③ For regional economies: It has formed "Yangtze River Delta/Guangdong-Hong Kong-Macao Greater Bay Area manufacturing and mass production" clusters; special actions in multiple provinces and cities (Beijing Wantai's mass production and hundred-billion-yuan cluster, Shanghai's 1 billion yuan fund) drive regional industrial chain upgrading and city cluster coordination, making embodied intelligence an "anchor industry" for regional industrial upgrading.

④ Dual ecosystem mechanism: It demonstrates a coordinated ecosystem of "policy procurement traction + top-level standards design"—in February 2026, the Ministry of Industry and Information Technology's Technical Committee for Standardization of Humanoid Robots and Embodied Intelligence released the "Standard System for Humanoid Robots and Embodied Intelligence (2026 Edition)," covering six major sections: basic commonalities, brain-inspired and intelligent computing, limbs and components, complete machines and systems, applications, and safety and ethics. It is the first top-level standards design covering the entire industrial chain and full life cycle, pushing the competitive dimension from the number of patents to "rule-defining power" and "governance in advance."

1.5.4 Complementarity, Convergence and Common Direction

(1) Complementarity.

The two paths are not opposed but complementary: the United States fills in the "body's" mass production and cost, and China fills in the "brain's" general capabilities and basic research, making the global ecosystem more complete.

(2) Convergence point.

A commercialization inflection point requires three conditions to hold simultaneously: usable algorithms, mass-producible hardware, and commercially viable factories. In 2025–2026, these three conditions came close to holding simultaneously for the first time; real commercial deployments (BMW–Figure, Tesla internal, Amazon–Digit) jointly validated PMF.

(3) Common direction.

Putting standards and governance "up front" is a consensus among all parties (to avoid repeating the regulatory lag seen in autonomous driving); cybersecurity has become the first area where genuine agreement has formed; talent (top-tier composite "algorithm–hardware–system" talent) is becoming a harder constraint than capital.

(4) Openness is inevitable, but it needs boundaries.

The divergence between the Chinese and U.S. paths (the United States in the "brain," China in the "body") makes openness in the embodied intelligence industry a necessity: first, structural complementarity—openness lowers cost curves on both sides; second, the data flywheel needs scale—open datasets and shared teleoperation accelerate global capability improvement; third, open-sourcing foundation models frees resources from "building brains" to "building bodies + finding scenarios."

However, the two countries are stuck at different positions in the industrial chain, and the struggle over supply chain security and the right to define rules requires "limited openness" at key layers. The four layers of meaning and boundaries of openness:

Table 1-5 The four-layer structure of "openness and boundaries" in China-U.S. embodied intelligence

In summary, openness in embodied intelligence is not unconditional globalization, but a competitive-cooperative landscape of "open-source as much as possible at the technology layer, proactive contestation at the standards layer, and holding the bottom line at the supply chain and market layers." This tension between "inevitable openness" and "limited boundaries" is precisely a true portrayal of the complementary convergence between China and the United States, and it also sets the basic framework for the evolution of the global industrial ecosystem.

2. Overall Trends and Outlook

The embodied intelligence industry has in 2025

-2026 officially crossed from proof of concept into commercial validation, and competition among cities has accordingly escalated from "grabbing projects and landing enterprises" to a "war of ecosystems"—a systemic contest over supply chains, computing power, scenarios, talent, and policy. Building on the aforementioned scoring framework, this chapter first presents several macro judgments on the direction of industrial evolution, then translates them into decision-making perspectives that city managers can directly leverage.

2.1 Macro judgments on industrial evolution

Judgment 1: The China-U.S. dual-core pattern will be hard to reverse in the short term, but the dimension of competition is moving upward.

The United States holds the innovation high ground in original large models, top-tier capital, and flagship enterprises, while China builds scale barriers through engineering capability and a complete industrial chain; together the two countries occupy seven of the global top ten cities. More critically, the focus of competition is shifting from "who can first produce a demonstration prototype" to "who can first push the brain, body, and scenarios down to a price point for scalable commercial use"; delivery capability and the speed of continuous iteration are replacing one-off demonstrations as the main basis for judging competitiveness.

Judgment 2: Europe is taking a differentiated path of "research + governance."

Zurich, London, and Munich maintain influence through top-tier research and a voice in AI rules; they do not pursue complete-machine scale, but instead position themselves in higher-value-added segments such as "standards, certification, and rules." This is not a challenge to the China-U.S. dual-core pattern, but rather a way to bypass head-on competition and point to a third path often obscured by the China-U.S. narrative: cities that do not pursue complete-machine scale can still be the "rule-makers" with the most say.

Judgment 3: The commercialization threshold is approaching, and scale deployment is beginning to materialize.

In 2026, humanoid and broadly defined embodied products are moving from demonstration to mass production; scenarios such as warehousing and logistics and industrial inspection have moved from factories and warehouses into scale deployment, and Morgan Stanley has twice raised its China shipment forecasts. The real meaning of the threshold lies in this: early-mover cities will irreversibly lock in talent density, latecomers' cost of catching up will rise steeply over time, and the window is narrowing.

Judgment 4: Industrial value is converging toward three layers: "brain–key hardware–scenario data." Profit and barriers are increasingly concentrated in three layers: embodied large models (the "brain" that determines the upper limit of intelligence), precision actuation and sensing hardware (the "body" that determines the lower limit of motion and perception), and high-frequency real-world scenario data (the "fuel" that determines iteration speed). Cities that try to take all of the stack often find it hard to build barriers at any single layer due to dispersed resources; precisely positioning at one of these layers, by contrast, makes it easier to establish irreplaceability.

Judgment 5: The industry will enter a period of differentiation and reshuffling within three to five years.

As the technology maturity curve passes its peak of expectations, capital and orders will concentrate on players with genuine delivery capability; complete-machine capacity lacking a differentiated core will face clearing out, and the ranking of city tiers will be rearranged accordingly.

Judgment 6: Soft infrastructure and compliance capability are becoming the threshold of the future.

As embodied systems enter scenarios such as elderly care and densely populated public spaces, functional safety, cyber, and ethical certification have changed from "bonus points" to "entry permits"; physical soft infrastructure such as public testing and validation platforms, open data and simulation, and computing power and green energy, together with institutional soft infrastructure such as data openness, standards certification, and governance rules, determine a city's short-term attractiveness to enterprises and long-term incubation capability. Whoever builds certification and standards capability first will hold the initiative in the next stage of competition.

2.2 Decision-making perspectives for city managers

Translating the above judgments into city actions yields five actionable orientations.

First, shift from "investment attraction" to "application pull." The asset a city can truly monopolize is its local scenarios—industrial inspection, warehousing and logistics, emergency rescue, elderly care, and so on. Rather than subsidizing enterprises to locate on the supply side, it is better to proactively open public-sector scenarios and use government procurement commitments to create demand pull, letting real orders screen the winners for the market. Scenario openness is itself the scarcest industrial policy tool a city has.

Second, diagnose and position yourself first, then choose a layer to occupy. Research-origination cities (such as San Francisco, Boston, and Zurich) should focus on the "brain" and talent density; manufacturing-industrial-chain cities (such as Shenzhen and Suzhou) should consolidate the "body" and supply chains; scenario-commercialization cities (such as Shanghai and Tokyo) should connect the brain

-body-scenario loop; governance-certification cities (such as Munich, Singapore, and London) should turn standards and certification into assets. Copying others' paths will only dilute already limited resources.

Third, replace "number of enterprises" with "ecosystem completeness" as the assessment yardstick. The real source of a city's competitiveness is the systemic combination of industrial-chain completeness, public training grounds, computing power and green energy, top talent density, and policy delivery. Only by shifting the KPIs for investment attraction and assessment from "number of enterprises that settle" to "completeness of ecosystem elements" can cities avoid fragmented stacking and match the reality that industrial competition has escalated to an "ecosystem war."

Fourth, guard against overcapacity and support "differentiation" rather than "homogenization." As the industry enters a period of divergence and reshuffling, city resources should be tilted toward players with differentiated core technologies, leading key-component companies, and entities with scenario-integration capabilities, rather than chasing homogenized complete-machine capacity. Positioning in key components (such as precision reducers, torque motors, and sensors) has higher strategic value than broad-based support for complete machines—the former builds barriers, while the latter easily becomes excess.

Fifth, plan ahead for soft infrastructure and rule-making voice. Physical soft infrastructure such as public testing and validation platforms, open data and simulation, and computing power and green energy, together with institutional soft infrastructure such as data openness, standards certification, and governance rules, are both the "infrastructure" for enterprises to land and the entry point for cities to participate in rule-making. Cities that take the lead in building certification and standards capabilities will not only be able to defend their local markets in the future but also export "entry thresholds" to the region and even the world, turning soft power into long-term influence.

In summary, city competition in embodied intelligence is no longer a sprint of single-point breakthroughs but a long-term contest over depth of understanding of industry trends and ecosystem-building capability. Only by first recognizing their own position in the value chain, following the convergence direction of "brain-body-scenario," and using scenario openness and soft infrastructure as levers can cities seize the initiative in this round of industrial restructuring.

English text is being prepared.