As 2026 passes the halfway mark, the humanoid robot industry has delivered a seemingly impressive report card.
Gan Xiaobin, Deputy Director of the Department of Science and Technology at the Ministry of Industry and Information Technology (MIIT), revealed at the July press conference for the World Artificial Intelligence Conference (WAIC) that the annual production of complete humanoid robots this year is expected to exceed 100,000 units. A five-fold leap from approximately 20,000 units in 2025 to 100,000 units is enough to draw the attention of any industry track. Ke Jixin, Vice Minister of MIIT, also pointed out that this surge in production sends a clear industrial signal: humanoid robots are accelerating their transition from technological exploration to standardized and scaled development.
However, behind the frenzy of production figures, the true landscape of the industry is far more complex than it appears on the surface. Within the nearly 100,000-square-meter exhibition hall of WAIC 2026, over 200 complete machine manufacturers competed side by side, ranging from bipedal to wheeled, from industrial to consumer, and from million-yuan to ten-thousand-yuan price tags. This competition has just begun, and the outcome remains far from clear.
01. Duopoly Leads the Pack, but Mass Production Does Not Equal Commercial Deployment
The first-half production rankings for humanoid robots are clear at a glance. Data from MIR DATABANK shows that Agibot and Unitree Robotics have established a "duopoly" pattern in the track. Agibot announced the rollout of 15,000 robots in early June, while Unitree completed the mass production of approximately 11,000 units in March.
Source: MIR DATABANK
Agibot's mass production trajectory can be described as textbook-level. With only 6 prototype units in 2023, it launched scaled mass production in August 2024, surpassing 1,000 units in January 2025 and reaching 5,000 units by year-end. By March 2026, it crossed the 10,000-unit milestone, breaking through 15,000 units in June. Yao Maoqing, a partner at Agibot, pointed out: "Mass production is the core foundation for embodied intelligence to step out of the laboratory and land in industries." Currently, Agibot has formed deployment solutions for seven real-world scenarios, covering production line loading and unloading, industrial handling, logistics sorting, tour guiding and shopping assistance, service retail, security patrol, and industrial and commercial cleaning.
Unitree is also accelerating. Owen Chen, the company's Asia-Pacific Director, revealed that this year's total shipments are expected to at least double the over 6,500 units in 2025, and might even approach a three-fold increase. To meet the demand, Unitree plans to expand its annual production capacity to 30,000 units. Until the first three quarters of 2025, over 70% of Unitree's humanoid robot revenue came from the scientific research and education sectors, with industrial applications accounting for only 9%. In other words, a large number of shipped robots did not truly "go to work in factories," but remained in laboratories and classrooms. This data reveals a subtle reality: the growth in production figures does not directly equate to the deep advancement of industrialization.
The MIR ranking also pointed out another noteworthy phenomenon: UBTECH, which ranked third in production, slipped to fourth in the shipments ranking. "High production but a drop in shipment ranking usually indicates inventory backlog or delayed delivery pace." The robots produced are not necessarily all delivered to customers. On June 30, UBTECH released the world's first mass-producible full-size ultra-bionic humanoid robot, U1, with cumulative orders exceeding 13,361 units, and the U1 Pro priced at 169,800 RMB. However, as a ToC consumer-grade product, these orders were not included in the industry's production statistics. The boom on the consumer end and the calmness on the industrial end form a stark contrast.
The temperature difference between production and shipments essentially reflects the most core contradiction in the humanoid robot industry today: manufacturing capacity has been initially established, but real user demand and a sustainable commercial closed loop are still under construction.
02. Differentiation: Five Companies, Five Survival States
Delving deeper into the industry chain provides a clearer view of this differentiation. At WAIC 2026, the author visited five humanoid robot companies with different backgrounds, and their situations almost outline the ecological map of the entire industry.
The National-Local Co-built Humanoid Robot Innovation Center is a representative of the pragmatic faction. Its humanoid robots have achieved stable commercial deployment in three types of scenarios: public guiding, security patrol, and industrial sorting and bin handling, with customers covering automation industries such as power and automotive. The planned total production capacity for 2026 is 5,000 units, with wheeled products priced at over 400,000 RMB and bipedal products at over 600,000 RMB. Notably, the Center frankly stated that in industrial scenarios, wheeled robots have lower costs, stronger stability, and better endurance, so wheeled shipments will be far higher than bipedal ones. This clear cognition of its own product form is not common in the current industry. Public information shows that the Center, together with Huawei, has built the country's first pilot training site for embodied intelligence. The pilot service platform created jointly with Shanghai Electric has been fully put into operation, opening up the whole chain of services including trial production, testing, and scenario verification to the entire industry, and has completed a special signing for the trial production of a thousand sets of humanoid robots with DroidUp. From the perspective of the industrial ecosystem, the Center's approach leans more towards the construction of platform-type infrastructure rather than the logic of a pure product company.
Matrix Intelligent Manufacturing takes the high-end bipedal route. Featuring a full-size bipedal design with dual dexterous hands, it uses linear joints instead of rotary joints for higher precision and stability. The third-generation general-purpose humanoid robot MATRIX-3, released during WAIC, stands 1.70 meters tall, weighs 65 kg, and is equipped with a 27-DOF cable-driven dexterous hand, capable of a single operation duration of 4 hours. Priced starting at 580,000 RMB, the planned production capacity for 2026 is 5,000 units, with cumulative orders of about 1,000 units. However, product functions are still in the adaptation and debugging stage—high price, small batch, and scenarios not fully finalized. The core logic of Matrix Intelligent Manufacturing is "scenario generalization," believing that the bipedal structure can cover the vast majority of spaces humans can enter without any environmental modification. This logic holds at the theoretical level, but the time window for commercialization and the cost curve remain huge unknowns. The company is currently still in the rhythm of "producing based on orders," which is still quite far from large-scale mass production.
Banxing Embodied Intelligence has taken a different path, adopting an ODM customization route, providing skeletal delivery of humanoid robots, allowing customers to customize the shell and brand themselves. Priced at 208,000 RMB, all core components are independently developed. It recently completed a new round of financing in the tens of millions of RMB range, with a monthly production capacity of dozens of units, and is expected to achieve an annual revenue of over 100 million RMB. This model reflects a certain embarrassment in the industry: customers who truly know what they want and have secondary development capabilities do not need to buy complete machines; while those who want to buy complete machines are often unsure of what they can do with them after purchase. Banxing's approach is equivalent to handing over the "last mile" to the customers to complete themselves, but this also means that its commercial ceiling depends on how many buyers in the market actually possess secondary development capabilities. From on-site exchanges, the company's attitude towards scenario adaptation is relatively pragmatic, frankly admitting that it will not promise customers "ready to use right out of the box," but realistically informs them of the product's actual capabilities and limitations.
Zhongjian Technology is still lingering in the concept stage. During WAIC, its subsidiary Hua Zhijian exhibited the concept humanoid robot ZERO, with R&D expenses in the first quarter increasing by 291.23% year-on-year. However, from what was learned on-site, ZERO is still in the prototype R&D stage, and specific application scenarios have not yet been clarified. Although R&D investment is huge, where the landing point for commercialization lies remains unknown. Zhongjian Technology previously judged the industry's enthusiasm with caution. In its view, there are still many realistic obstacles to the deployment of humanoid robots, and the actual progress of the industry may not be as optimistic as market expectations. This conservative posture reflects from one side that the internal attitude towards the "mass production frenzy" is not unified within the industry.
Magic Atom, on the other hand, chose to penetrate vertical scenarios. Its traffic duty robot "Xiaomai" has been certified and on duty to perform traffic guidance tasks at the Wuxi Marathon this April. During WAIC, it also released new products such as the flagship full-size humanoid robot MagicBot X1 and the industrial wheeled humanoid robot MagicBot D1. Magic Atom is the most focused on scenarios and has the clearest customers among the five companies. Although the market size of the vertical scenario of traffic duty is limited, it excels in real demand, clear willingness to pay, and a clear deployment path. In addition, beyond the traffic scenario, the company is also extending to industrial and service scenarios, attempting to replicate the engineering capabilities accumulated in vertical scenarios to broader application fields.
From those already deployed with revenue, to the prototype stage, and then to ODM customization, the five companies with five different situations outline an industry landscape that is far more complex than "all going well."
03. The Form Myth: Bipedal, Wheeled, or Robot Dog?
Is "humanoid" necessary? This is probably the most fundamental yet most divisive question in the current industry.
The core logic of the bipedal faction is "environmental compatibility." Matrix Intelligent Manufacturing believes that the current human living environment is designed on a human scale. The height of stairs, the width of corridors, the horizontal position of operating tables—all infrastructure is planned according to human body dimensions. The bipedal structure can seamlessly integrate into this existing world without requiring any modification to the environment. In contrast, wheeled solutions are more efficient on flat ground but are limited when encountering stairs, thresholds, and uneven surfaces; quadruped solutions have advantages in complex terrain adaptability, but in the daily human environment, two legs are already the "optimal solution" verified by thousands of years of evolution.
But the perspective of the pragmatic faction is completely different. The structural stability of robot dogs is significantly superior to bipedal ones, with a lower center of gravity, a smaller probability of falling, and stronger anti-interference capabilities. Under current engineering levels, these characteristics mean lower failure rates and higher availability. The shipment data from the National-Local Co-built Center also confirms this: wheeled robots are more widely used in industrial scenarios, and this year's wheeled shipments are far higher than bipedal ones. A more straightforward way to put it is that the core value of humanoid robots in commercial scenarios lies in satisfying emotional value—such as guiding, reception, and performance, while in industrial scenarios, the solution of an AGV chassis plus a robotic arm is already mature and efficient enough. In 2025, quadruped robot shipments exceeded 44,000 units, and in the first half of this year, it was indeed quadruped robots and wheeled dual-arm robots that were the first to "go on duty" on a large scale.
These two logics are not necessarily opposed, but they point to completely different time dimensions. Bipedal is a long-term vision, premised on motion control algorithms being mature enough, hardware costs being low enough, and battery energy density being high enough. Wheeled and quadruped are currently available solutions, with higher technology readiness levels, lower costs, and a greater likelihood of customers being willing to pay. The question is, how many companies can survive to the "long term"? If cash flow cannot last through the next two or three years, even the most perfect long-term vision is meaningless.
From a global perspective, the competition between different technological routes continues. Boston Dynamics' Atlas has demonstrated astonishing mobility after switching to electric, but its cost structure remains a mystery. Tesla's Optimus mass production targets have been repeatedly delayed, which to some extent indicates that the engineering difficulty of bipedal humanoids far exceeds expectations. Domestically, Unitree and Agibot have run through scaled shipments on the consumer and scientific research ends, but penetration on the industrial end remains limited. A practitioner's judgment is quite direct: for bipedal humanoids to gain a firm foothold in industrial scenarios, at least two core issues need to be solved: endurance and reliability, and neither has a satisfactory answer at present.
04. Capital is Still Pouring In, but the Money Has Changed Direction
In the first half of 2026, the total financing in the embodied intelligence track reached 104.1 billion RMB, nearly double the total for the whole year of 2025. Unitree Robotics' STAR Market IPO registration has been approved,LimX Dynamics completed a Pre-IPO financing round of nearly 200 million USD, and the listing applications of Leju Intelligence and DeepRobotics have been accepted one after another. A wave of listings is accelerating.
However, structural changes have occurred in the flow of capital. Over half of the money has rushed to "brain faction" companies, with robot body manufacturers receiving less than 20%. Capital is no longer chasing bodies that "can walk and jump," but instead betting on brains that "can think and make decisions." VLA models swallowed 42% of the funds in one go, world models accounted for 27%, and data infrastructure accounted for 20%. This flow of funds itself is a barometer of the industry's bottleneck. What truly chokes the industry is not hardware manufacturing capacity, but the level of intelligence that enables robots to understand the world and make autonomous decisions.
This shift points to the most fundamental bottleneck in the industry. VLA (Vision-Language-Action model) is universally regarded as the standard component of the embodied intelligence brain, but to achieve general autonomous capabilities, at least tens of millions of hours of high-quality real interactive data are needed. The reality is that as of early 2026, there are only 500,000 hours of compliant and available robot data globally, with a gap of over 99%. Currently, most humanoid robots are far from achieving autonomous consciousness, and at the practical operation level, they essentially still rely on manual remote control or preset programs. It can be said that the humanoid robot industry is currently facing the embarrassment of "having a body but no brain." Motion control (the cerebellum) is already quite mature, but perception and decision-making (the brain) are still at a rather primitive stage.
Batteries are another physical hard flaw. The limited space in the body makes it difficult to equip large-capacity batteries, and the endurance of existing products is generally between 1 to 3 hours. A practitioner calculated: for a 70 kg humanoid robot to merely maintain a standing posture without falling, the hip and knee joints need to continuously output torque to fight gravity, consuming 200 to 300 watts of power. It can run for four or five hours under fine-tuned laboratory conditions, but in real scenarios, lasting one or two hours is already quite good. The pace of improvement in battery energy density basically follows the laws of physics, making a breakthrough in the short term impossible. This means that the endurance issue will be a hard constraint for bipedal humanoids for the next few years.
05. The Elimination Race Has Already Begun
In February 2026, China's first "Humanoid Robot and Embodied Intelligence Standard System (2026 Edition)" covering the entire industrial chain of humanoid robots was officially released. In June, the MIIT and the State-owned Assets Supervision and Administration Commission (SASAC) jointly launched the annual real-scene practical training special action, clearly stating that application verification and regular deployment will be completed in a batch of representative scenarios by the end of the year. Ke Jixin, Vice Minister of MIIT, judged that the next two to three years could be a critical point, or even a new starting point, for the development of the humanoid robot industry.
Policies are being intensified, but the market's elimination logic is even more ruthless. Xie Shaofeng, Director of the Standardization Technical Committee for Humanoid Robots and Embodied Intelligence under the MIIT, also stated at the recent Humanoid Robot Standard Week that the current industry is "blooming in all directions yet fighting each on their own", with some companies still stuck in the prototype and demonstration stages, and the commercialization path remains unclear. A report from Qixinbao shows that there are over 680 domestic humanoid robot complete machine manufacturers, while the threshold for making the MIR ranking means that over 99% of the players are still just running alongside.
The judgments of multiple practitioners are highly consistent: the industry will experience a drastic survival of the fittest over the next three years, and a large number of manufacturers lacking core competitiveness will be washed out. Currently, the industry generally stays at the demonstration-level function stage, and the deployment capability that can truly stand the test of practice is the key yardstick for measuring enterprise value. A practitioner made an analogy: right now, everyone is just doing fancy but impractical kung fu, just showing off moves; the real kung fu can only be seen in actual combat. This sentence points precisely to the biggest bubble in the industry today: a large number of companies use prototypes and demonstrations to attract capital and attention, but have not truly answered the core question of "what exactly can my robot do."
Zhu Keli, founding dean of the Guoyan New Economy Research Institute, pointed out that at this stage, most humanoid robots are still mainly delivered as customized projects, with high unit prices, and the commercial cost-effectiveness of scaled procurement has not yet emerged. If a replicable deployment model cannot be quickly formed, the industry is prone to falling into a vicious cycle of "high cost, low sales." Zhao Mingguo, a professor in the Department of Automation at Tsinghua University, gave a precise prediction in January this year: "2026 may be the year of screening for application scenarios."
Production surging to 100,000 units does not mean 100,000 robots are working in real positions. For Unitree and Agibot, it is a price war and shipment volume competition on the consumer end; for the National-Local Co-built Center, it is the steady penetration of government service scenarios; for Matrix Intelligent Manufacturing, it is a long-term bet on technological routes; and for Zhongjian Technology, it is a question that has no answer yet.
As one practitioner said, it is meaningless to just make robots that can dance. When the spotlights go out and the exhibitions end, the real question is, who can stand firm in real positions? The answer may take a few more years. But the direction is already very clear: humanoid robots that cannot do real work will eventually be eliminated by the market.