Three signals landed within the same fortnight, and together they describe the structural question the humanoid industry will carry for the next decade. On September 7, Reuters published an investigation by Eduardo Baptista documenting how China’s defense establishment is systematically pulling humanoid technology toward military use. Days earlier, reporting on the Chinese sector showed startups earning measurable revenue in airports, hotels, banks and logistics centers rather than staging demos. And the field’s flagship academic meeting, IEEE Humanoids 2026, has rebuilt its December program around a theme nobody would have proposed in 2022: labor conditions, safety certification and economic impact.
One machine, three customers. The factory wants throughput. The service counter wants reliability around people. The military wants expendability. All three are now drawing from the same supply chain, the same model architectures and increasingly the same training data. The third customer just became documentable, and the record deserves a closer read than the headline “China readies humanoid robots for combat” allows, because what the documents actually show is both more mundane and more consequential than the framing suggests.
What the Reuters Record Actually Establishes
The investigation is not built on speculation. Reuters reviewed more than 100 Chinese military procurement notices, academic studies, patents, official publications, government records and defense-company materials, and the timeline it reconstructs is precise.
Two days after the World Humanoid Robot Games ended in Beijing on August 26, the PLA’s official newspaper drew its conclusions from the spectacle. The PLA Daily called for researchers to accelerate the transfer of cutting-edge technologies from laboratories to military training grounds, explicitly framing them as robotic “combatants.” That editorial was not an isolated flare. A July 2025 PLA Daily article had already argued that humanoid missions could evolve “from supporting combat toward primary combat,” and discussed the conditions under which a robot might be authorized to fire on a living target that had been verified by a human.
The institutional paper trail runs deeper than commentary. In August 2025, researchers at the National University of Defense Technology’s test center in Xi’an published a study modeling an urban assault force: six combat robots, a mix of humanoids, robot dogs and unmanned vehicles, split between two assault teams working with ground troops to clear an enemy-occupied building floor by floor. The study projected the required equipment could be available within five to 10 years. Notably, it did not specify rules of engagement, weapons or how civilian encounters would be handled. A June 2025 NUDT university competition went further into scenario design, with an event simulating “infiltration behind enemy lines” requiring robots to enter an area, map it and locate targets, and another envisioning identity checks and security patrols at military bases.
Procurement followed research. In May 2025, the PLA issued a tender to purchase a humanoid robot, including installation and technical training. Four months later, a procurement notice sought an “embodied humanoid robot intelligent perception and dexterous operation system.” In June this year, the PLA budgeted roughly $300,000 for a system to collect and label camera, radar and motion data for humanoids, including information about terrain they could traverse. Reuters could not determine whether the tenders were filled.
On the industry side, Norinco, the state-owned defense conglomerate, published materials in August describing its full-size Fuxi humanoid as capable of sentry duty, all-weather reconnaissance, intelligent patrol and dangerous tasks, paired with a teleoperation system that lets a human operate the machine remotely. And in December, the PLA’s Eastern Theater Command released an AI-generated montage showing military robots, humanoids among them, overwhelming Taiwan’s defenses in a hypothetical future conflict.
Read as a ledger, this is demand-side evidence: an army’s research university hosting competitions around infiltration scenarios, procurement officers issuing tenders, a defense prime shipping product marketing. None of it is a vendor press release.
What the Record Does Not Show
The same investigation is scrupulous about the gap between interest and capability, and that gap is the story’s most useful analytical content.
Reuters found no evidence that China has deployed an armed humanoid with an operational PLA unit. The robots remain energy-intensive and unreliable outside controlled demonstrations, a point Reuters put to Aaron Johnson, a mechanical engineering professor at Carnegie Mellon, who noted that battlefield environments challenge robots in unpredictable ways. Malcolm Davis of the Australian Strategic Policy Institute dismissed the near-term practicality entirely: not useful today, but “in five to 10 years, they could very well be.”
Juo-Min Chou of Taiwan’s Institute for National Defense and Security Research made the economic point that matters most: for reconnaissance and logistics missions, quadruped, tracked or wheeled robots are generally cheaper and easier to deploy at scale. The humanoid’s theoretical advantage is the human-shaped niche, buildings, tunnels and ship interiors, where two hands and two feet combine mobility, climbing and manipulation. The battlefield use case is a subset of the general-purpose use case, not a shortcut around it.
Even the most bullish military researchers concede the dependency chain. Wang Yonghua of China’s Academy of Military Sciences wrote in a November commentary that humanoids face unresolved problems in movement, perception and intelligence, and that technical breakthroughs, lower manufacturing costs and industrial-scale production would be required before the machines become widespread military systems. If those conditions are met, he wrote, “humanoid robots will stream onto the battlefield.”
That is a conditional sentence, and every clause in the condition is a civilian-market problem. Cost curves bend in factories. Scale comes from logistics contracts. Perception and intelligence improve on commercial data. The military timeline is downstream of the commercial one, which is exactly why the PLA’s interest focuses on perception, manipulation and training data rather than on buying bodies.
The $300,000 Tender That Should Get Everyone’s Attention
Buried in the procurement record is the item with the widest implications. The June 2026 PLA budget line, about $300,000 for a system to collect and label camera, radar and motion data including terrain traversability, is a small purchase with a large signature.
It describes precisely the product the commercial market has been organizing itself to sell: labeled embodied data. The VLA data economy, teleoperation traces, egocentric video, robot execution logs, has become one of the hottest commodity markets in robotics, and a military labeling pipeline is dual-use by construction. A company building data infrastructure for factory robots does not get to choose whether the same pipeline feeds a weapons program; the data does not know its customer. This is the pattern to watch, more than any single robot demo, because it is how dual-use technology transfer actually happens: not through dramatic hardware transfers but through shared inputs, shared benchmarks and shared tooling.
There is a supply-side mirror to this. Chinese manufacturers accounted for roughly 95 percent of global humanoid shipments in 2025, according to BofA Global Research cited by Reuters. Any military humanoid program anywhere, including the US Army’s own competition for “militarized humanoid capabilities” launched last year, with up to 10 finalists testing alongside American military experts and as much as $1.25 million in follow-on contracts, inherits an industrial base whose center of gravity is Chinese. Reuters separately reported that Unitree’s best-selling robot dogs were based on innovations originally financed by the US military. The technology flows are already bidirectional and nobody’s supply chain is clean.
Defense Was Already in the Funding Stack
The uncomfortable fact for anyone who imagines a bright line between civilian research and military applications is that the line has never existed in this field. The funding credits of recent landmark papers read like a defense portfolio. BeyondMimic, the UC Berkeley and Stanford framework that brought whole-body humanoid control to general-purpose, data-driven training and appeared in Science Robotics last month, lists DARPA’s TIAMAT program among its funders. The locomotion layer that is now rapidly saturating, the layer that makes a sprinting, kicking, cartwheeling robot almost routine, was subsidized on exactly this dual-use logic.
UCLA’s Dennis Hong gave Reuters the expendability case in its cleanest form: going where we do not want humans to go is one of the most compelling reasons to develop robots at all, and a robot that can be lost instead of a person is a robot that can be spent. That argument predates every company currently shipping humanoids, and it will outlast them.
The difference now is scale on the civilian side. When DARPA funded the DARPA Robotics Challenge in 2013, there was no industrial humanoid economy to militarize. Today the same agency’s research money lands in a field with factory deployments, service contracts and public markets, which means each technical advance books to three accounts simultaneously, whether anyone intends it or not.
The Civilian Counterweight: Where the Money Actually Is
Against the military pull, consider what the commercial record from the same weeks shows, because it is the load-bearing wall of the entire dual-use question. Inside AI’s survey of the Chinese sector documents a market that has quietly diversified beyond demo videos into revenue.
Unitree, the company whose UnifoLM-X2 world model powered last week’s fully autonomous combat demonstration video, reported RMB 1.15 billion in first-half revenue, up 48.5 percent year over year, with cumulative production of about 18,000 bipedal humanoids. Its August listing on Shanghai’s STAR Market is a cautionary tale in itself: shares opened at RMB 1,100 against a RMB 150.80 offering price, briefly valuing the company at RMB 444.9 billion, and then lost more than half their value in two weeks. The market is willing to fund the boom and willing to discipline it.
The deployment texture is where the real signal lives. Hefei-based Zerith has wheeled humanoids working commercial shifts at Hefei Xinqiao International Airport and Olé supermarkets, with pilot robot stores in Shanghai whose best location clears RMB 20,000 a week. Beijing’s RobotEra has robots working with China Post and SF Express across more than 10 logistics centers and orders exceeding RMB 500 million, on the strength of dexterous-hand manipulation work. Keenon, which has shipped over 100,000 robots to more than 70 countries, runs eight machines at a Shanghai Shangri-La handling greetings, luggage, deliveries and cleaning, with the hotel reporting roughly 30 percent lower demand for repetitive labor, 40 percent faster service response and 35 percent shorter guest waits. Digital Huaxia has humanoids in more than 20 ICBC bank branches with first-year orders at the RMB 100 million level.
Every one of these numbers is a civilian number. The unit economics that Wang Yonghua’s condition requires, lower costs through scale, industrial production, better perception, are being built by hotels, banks, post offices and supermarkets. The military inherits whatever this economy produces. It does not generate it.
December’s Governance Fork
The academic field, for its part, is not pretending the ambiguity away. IEEE Humanoids 2026, the 25th edition of the field’s flagship conference, runs December 6 to 9 in Santa Clara under the theme “Humanoids and the Future of Work,” with more than 1,200 expected attendees from over 70 countries and a program that puts safety certification, worker impact and deployment governance alongside the technical sessions.
The most consequential item on that agenda may be the Loco-Manipulation Challenge, which will run 10 to 15 teams through a task circuit spanning mobility, manipulation and dexterity, and reasoning. Its scoring design is quietly radical: it rewards autonomy but still grants points for teleoperated performance, an explicit acknowledgment that partial autonomy remains commercially useful and that the field needs public measurements of where autonomy actually breaks down.
That measurement problem is the same one Norinco’s teleoperation system answers in the military domain. The honest state of the art, on both sides of the Pacific, is machines that can execute impressively inside controlled conditions and degrade unpredictably outside them. A public benchmark that documents the autonomy boundary serves civilian procurement teams and arms-control analysts alike, because “fully autonomous” in a marketing video and “autonomous under adversarial conditions” are different claims separated by exactly the distance the challenge proposes to measure.
The normative layer is thin but not empty. The International Committee of the Red Cross has cautioned that autonomous systems may struggle to interpret signs of surrender, judgments that depend on context, conduct and intent. The US military requires legal review and realistic testing for autonomous weapons, with commanders retaining judgment over use of force. China’s own defense ministry said in March that AI-powered weapons should remain under human control. The July 2025 PLA Daily article’s framing, a robot firing on a target verified by a human, sits inside that consensus language while quietly outsourcing the hardest part, verification, to the machine.
What to Watch
Four markers will tell you whether the dual-use ledger is filling or staying thin. First, whether Chinese procurement notices shift from training systems and data infrastructure toward unit purchases, which would signal capability confidence. Second, whether US and allied export-control lists extend beyond semiconductors into humanoid components, actuators and embodied training data, the natural next front if the 95 percent shipment statistic hardens into policy. Third, the December loco-manipulation results, the first widely visible public measurement of the autonomy boundary across academic and commercial platforms. Fourth, whether Western humanoid vendors begin disclosing defense revenue at all, a line item that is currently invisible in an industry that sells itself almost entirely on labor economics.
The sector spent 2024 through 2026 telling a single story about labor shortages and factory floors. That story is still true, and it is still where the money is. But the Reuters record shows a second buyer has entered the room, patient, well-documented and institutionally patient about timelines, willing to wait the five to 10 years that experts on three continents independently cite as the realistic window. Every humanoid company now holds a defense option it never asked to purchase. Whether the industry’s safety, certification and governance work matures faster than the military pull is the competition that actually matters, and it has no benchmark yet.