Steel Flock: How China is building a robot army and what it's actually testing
In April 2025, the first in stories A half-marathon of humanoid robots. Twenty-one participants started, and only six made it to the finish. The rest fell, overheated, and lost their balance on the turns: a short video of a humanoid toppling over went viral on social media faster than the official press release announcing the winner. Five months later, the four-legged Unitree robots were parading down Chang'an Avenue, past the stands where Xi Jinping stood. Between these two images lies the entirety of modern Chinese combat robotics: a large industrial base, a high-profile public display, and a vague distinction between showmanship and combat effectiveness.
Goliaths and Teletanks: A Brief History of the Combat Robot
The first mass-produced combat robots appeared in the USSR in the 1930s. Ostekhbyuro assembled the TT-26 teletanks and TU-26 control vehicles based on the light T-26. These vehicles were radio-controlled and armed with a machine gun and flamethrower. Their use in the Winter War of 1939–1940 proved limited: communications were disrupted, and the control vehicle crew tank I lost sight of my charge in the first ravine. But the principle was clear: car in front, person behind.
A teletank is in the foreground, with a command tank in the background. The teletank is easily identified by the distinctive "ears" of the whip antennas on its turret.
The Germans took a different approach during World War II. The Goliath self-propelled mine was a tiny tracked vehicle with an electric motor or a two-stroke gasoline engine, 60–75 kg of explosives, and wire control. More than seven thousand were produced between 1942 and 1945; the wire was easily damaged by small arms fire, but the design itself survived as an idea into modern mine-clearing robots.
Goliath self-propelled mine
After the war, the field faded into obscurity for a long time. The current surge in activity was spurred by two American programs in the 2000s. The DARPA Grand Challenge series of self-driving car races from 2004 to 2007 brought autonomous navigation algorithms to practical use. And the BigDog, a quadruped with an internal combustion engine, unveiled by Boston Dynamics in 2005, became the technological ancestor of all current quadruped platforms, including those produced by Unitree today. China's "robot army" has emerged from two waves: American research programs and its own industrial base of the 2010s.
The autonomous transport robot was a participant in the DARPA Grand Challenge.
From Imports to Production: How China Built an Industrial Base
Ten years ago, approximately three-quarters of China's demand for industrial robots was met by imports from Japan, Germany, and Switzerland. By 2024, the situation had reversed: according to the Center for Strategic and International Studies (CSIS) (officially designated an "undesirable organization" in Russia), 57% of industrial robots installed in the country were made domestically. That same year, 295,000 new industrial machines were installed in China, more than the rest of the world combined. The density is 166 robots per 10,000 workers; for comparison, this figure is 1220 in South Korea and 818 in Singapore. Korea and Singapore remain ahead, but China's growth rate is different: industrial robot production in 2024 increased by 14% year-on-year, and by 28% in 2025.
Foreign expansion is following the domestic market. China's share of global industrial robot exports has grown, according to CSIS estimates, from 5,9% in 2020 to 16,7% in 2024. The gap in the humanoid robotics market is even more pronounced: according to the International Federation of Robotics, there are approximately 80 companies working on humanoid robotics in China, an order of magnitude more than in the US. Production capacity is also visible. In January 2026, China's first humanoid production line was launched in Foshan (Guangdong Province), with a projected capacity of 10,000 units per year. AgiBot's 65,000-square-meter factory is designed to produce over 5000 A2 robots annually; by the end of 2025, it had assembled approximately 2,000 (less than half of the stated capacity, which is normal for the first year). The base version of the Unitree quadruped robot sells for approximately $6000; by comparison, Elon Musk has at various times quoted a launch price for the Tesla Optimus between $20,000 and $30,000.
Upon closer inspection, this industry only appears seamless. There are plenty of gaps. High-precision joint gearboxes, both harmonic and cycloidal, are largely sourced from Japan: Harmonic Drive Systems and Nabtesco hold the majority of the global market, and Chinese manufacturers haven't yet learned to outshine them with their equivalents. NVIDIA supplies GPU-level hardware for neural network processing; EngineAI also uses AGX Orin modules in its humanoids. Lithography for producing in-house chips is limited by access to ASML equipment. As long as these components are supplied reliably, the industrial machine operates; a serious supply disruption will reveal the extent of localization.
An important nuance is the industry's direction. In the USSR and the USA in the 20th century, military robotics followed suit: first the Ministry of Defense program, then conversion (roughly speaking, first the tank, then the tractor). In modern China, it's the opposite: the mass industrial and consumer segments lead, and military applications derive ready-made products from them. The strength is obvious: low production costs and a well-established supply chain. Less obvious is the inverse relationship: the program relies on commercial demand, which can falter.
Wolves, humanoids, and the "mother": what the tests showed
Four-legged robots. The most recognizable Chinese combat vehicle is the Unitree quadruped robot in its military modification. In Chinese terminology, it's called "wolf," or 战狼. According to the developer and exhibition data, the latest generation's specifications include: speed of up to 15 km/h, payload of up to 25 kg, ability to overcome obstacles up to 30 cm high even when fully loaded, and 12 degrees of freedom in the joints. On board: 4D LiDAR with a wide field of view, visible and infrared cameras, and a Huawei Ascend 310B processor for real-time data processing. Materials from the Foundation for Defense of Democracies describe the role specialization of the "pack": reconnaissance vehicles with additional sensors, attack vehicles with small arms weapons or a grenade launcher, ammunition transport. One operator controls the system via a tactical glove, voice commands, and a joystick on the rifle's fore-end.
The exact performance characteristics of the military versions have not been published in open sources: what is shown at parades and what is actually in service with the People's Liberation Army of China (PLA) may differ. Not a single combat use of the "Wolves" in a real conflict as of May 2026 has been documented in open sources. Energy is a separate issue: the Unitree's battery life remains a weak point of the platform, especially in freezing temperatures. Overheating during the Beijing Half Marathon is a testament to the same physics that will persist in the field.
Humanoids. The EngineAI T800 platform, showcased at the World Robotics Conference in Beijing. According to the exhibit and the developer's press materials, it stands 1,75 meters tall, weighs 75 kg, and has 29 degrees of freedom throughout its body, including 7 in each hand. The body— aviation Aluminum, torque in high-torque joints up to 450 Nm. Sensors include 360-degree LiDAR and stereo cameras. Computing power includes an Intel N97 as the base module and NVIDIA AGX Orin for neural network tasks, with a claimed performance of 275 TOPS (trillion operations per second). In terms of computing power, this is comparable to L3 automotive driver assistance systems, meaning the hardware is sufficient for simultaneous processing of a video stream, a lidar point cloud, and thermal images in real time.
Aviation mother. In June 2025, the Jiu Tian SS-UAV unmanned aerial vehicle was unveiled. It has a 25-meter wingspan and two internal bays for 100–150 small UAVs: loitering munitions, strike drones or decoys. Conceptually, this is the Chinese answer to the American Skyborg and XQ-58 Valkyrie programs, a carrier of "faithful wingmen," or loyal wingman. The idea is closer to the Russian S-70 Okhotnik paired with the Su-57 than to a classic attack UAV like the MQ-9. And historically, it is a great-great-grandson of the Soviet "Zveno" project by engineer Vakhmistrov: in the 1930s, up to five fighters were attached to the TB-3 heavy bomber, and this combination managed to fight in the summer of 1941; strikes on the Carol I Bridge over the Danube at Cernavoda and the oil depot in Constanta are known. The aircraft mothership is almost a century old. Only the size of the attached aircraft and the amount of control it has varies. The layout is completed by unmanned L-30 boats with a speed of up to 65 km/h and Guangjian laser systems for combating small drones; It is no longer the “army of robots” itself, but the environment in which it must operate.
Unitree and the "Little Dragons": How the State Comes to the Company
A disclaimer regarding sources. Most of the public analysis on Chinese military robotics comes from American and Western centers: CSIS, FDD, Georgetown University's Center for Security and Emerging Technologies (CSET), and Morgan Stanley. Chinese public analysis does exist (materials from the Chinese Academy of Social Sciences, the Jiefangjun Bao, and publications from the National University of Defense Technology), but it is less accessible and emphasizes differently: less about the threat and more about doctrinal normalization. This doesn't invalidate Western assessments; it simply means that the picture in this text is painted from one side of the ocean.
Unitree Robotics was founded in 2016 by engineer Wang Xingxing, who left DJI after just a few months to develop affordable quadruped robots. Initially, they were intended for labs, film studios, and shows. In October 2022, Unitree, along with Boston Dynamics, Agility Robotics, and three other companies, signed an open letter publicly committing not to arm their machines and to verifying buyer intentions. Two years later, in 2024, Unitree B1-series quadruped robots with assault rifles strapped to their backs were filmed on state television during the joint Chinese-Cambodia Golden Dragon 2024 military exercises. In August 2025, the company issued an official statement: Unitree is a civilian company; all military-grade modifications are made by third parties. This statement is difficult to independently verify: there is no public registry of "third parties," no audit of the assembly chain has been conducted, and specific integrators of the weapon systems are not named in public documents. In September of that year, these same vehicles paraded through Chang'an. And in February 2026, the US Department of Defense added Unitree to its list of Chinese military companies, citing government support and ties under the civil-military merger program. The furor surrounding this decision wasn't solely due to military considerations: by that time, Unitree quadcopters were widely sold worldwide, including to research labs in the US and EU.
A video of a robot carrying a rifle on its back is an effect. More interesting is the incentives behind a civilian company's move into the military sphere. China's Ministry of Industry and Information Technology has designated Unitree as a "small giant" under the Zhuan Jing Te Xin program, which provides preferential tax treatment, subsidized lending, and preferential access to government procurement. In January 2025, the Hangzhou administration designated Unitree as one of the "six little dragons": according to media reports, the city allocated approximately 16% of its industrial budget to these six. The Hangzhou High-Tech Zone itself, where Unitree is based, is, according to a War on the Rocks report, a designated center for civil-military fusion, and the local development bureau has an institutional mandate to coordinate "civilian participation projects in the military sphere."
The connection between Unitree and the military program is based on status, benefits, and access to government procurement, rather than a classic defense contract. According to a Georgetown University study, between January 2023 and December 2024, the PLA awarded 2857 contracts related to artificial intelligence; 1560 different organizations received at least one. Of the 338 that won two or more, approximately three-quarters were companies without state ownership. The commercial sector plays a significantly larger role in this program than it did in the Soviet or American models of the 20th century. This is its main distinction.
Intellectualization: Algorithms are more important than hardware
Within the PLA, the development of combat robotics fits into a doctrinal triad: mechanization, informatization, and intellectualization (智能化). Mechanization was officially declared complete by 2020. Informatization refers to the network-centricity of the 2000s–2010s, familiar to Western armies as well. Intelligentization has formally been a priority since 2019; it encompasses autonomous systems, neural networks, big data processing, and related AI. What this means in terms of specific KPIs (key performance indicators) is a separate question, and there is no answer in public documents.
In engineering terms, this primarily means working on local coordination of drones and ground robots without constant radio communication, i.e., resistance to electronic warfare. In February 2025, researchers from the PLA Army Engineering University and the National University of Defense Technology published a paper on self-organizing networks of attack drones capable of performing missions in jammed environments. The idea itself has long been discussed: a single drone is dumb, while a network of drones in a local network must be smart. This is the entire modern concept of swarm warfare in a nutshell. Beyond that, the testing grounds begin, and beyond that, the algorithms tend to get more interesting. During the Joint Sword-2024B exercises in October 2024, among other incidents, a group launch of attack drones with automatic target redistribution was revealed; there are no public analyses of how independently the algorithm operated and where the operator intervened. But this was one of the publicly shown instances in which the Chinese side demonstrated the swarm component, rather than individual platforms.
Here, too, is a rethinking of the so-called OODA loop, the "observe-orient-decision-act" cycle. The concept was formulated by American fighter pilot and military theorist John Boyd. In the 1950s, he flew the F-86 in Korea and tried to understand why American pilots won air battles against the MiG-15, which nominally had an advantage in climb rate, ceiling, and armament. The answer he arrived at in the 1970s was counterintuitive: it turned out to be a matter of decision-making speed, not aircraft performance. The winner is the one whose "see-understand-decide-act" cycle is shorter. Modern Chinese doctrine is based on a simple premise: a machine completes the OODA loop faster than a human. The logic is then obvious: some decisions are delegated to the algorithm. On paper. In practice, this comes down to the rules of engagement, which the PLA has not published publicly. China is participating in negotiations under the UN Convention on Inhumane Weapons, but its official position allows for the development of systems that meet some of the lethal autonomy criteria, leaving a wide margin of safety.
The declared milestone is 2027, the centenary of the PLA's founding; in Xi Jinping's speeches, this date is cited as the deadline for achieving combat readiness for a new type of army. What exactly is meant by "readiness" is vague, and Western analysts interpret it differently. By 2035, modernization is planned to be "basically completed," and by 2049, to achieve a "world-class army." These are political declarations, not confirmed combat readiness figures; to take them for a timetable is to confuse goal-setting with results.
Parallel programs and where the boundary lies
In the US, the emphasis is on the quality of individual platforms and network integration: Boston Dynamics with Atlas, Figure AI, Agility Robotics, Tesla with Optimus. In August 2023, the Pentagon announced the Replicator program: a plan to deploy "several thousand" autonomous systems across "all domains" (air, sea, land) with targeted funding of approximately $1 billion for the first two iterations. The JADC2 (Joint All-Domain Command and Control) program, a network-centric superstructure that integrates sensors, platforms, and decision-making systems from all branches of the military into a single system, is underway in parallel; the US Department of Defense has requested approximately $4 billion for its development between 2022 and 2024. According to Morgan Stanley estimates, Chinese companies have produced more than twice as many humanoids as American companies since 2022, and have filed three times as many patent applications with the word "humanoid," but the US still remains ahead in terms of consolidated defense orders for autonomous systems and networked control.
In China, the logic is different: Unitree, EngineAI, XPeng, AgiBot—mass production, low cost, a ready-made supply chain. Europe—advanced programs in the absence of a comparable industrial base.
Calling this a lag between one and the other is like measuring the same thing with different yardsticks. The strategies are simply different. The Americans have a polished unit and a network circuit on top of disparate platforms. The Chinese have mass, in which losses due to failures are compensated by quantity. With an important caveat: "mass compensates for failures" works for expendable systems like kamikaze attack drones or for saturation missions. Defense, where the statistics favor the attacker. For platforms that are destined to return (urban reconnaissance, walking vehicles, humanoids with complex motor skills), mass production of cheap vehicles is no substitute for reliability. The math is different there. Which of the two strategies will prove more resilient in real-world deployment is an open question; neither side has yet provided a definitive answer.
In short: the program already has an industrial base. The doctrinal framework is still being developed. There's no combat experience whatsoever. The distance between a humanoid falling at the Beijing Half Marathon and a four-legged robot marching in parade formation is still greater than it seems from the impressive videos, but judging by the pace of factory construction in Foshan and the density of PLA contracts with civilian suppliers, it's closing. Somewhere in the middle. There's nothing more precise to say today, and anyone who talks about it is making things up.
Information