China Built 19,000 of the World's Humanoid Robots in H1, IDC Data Shows

IDC also raised its forecast for global humanoid shipments in 2030 to more than 750,000 units, TechNode reported. The same figures put global shipments for the half at nearly 25,000 units, up 432.1% year over year. TechNode credits IT Home, in Chinese, for the underlying numbers.

That is a shipment story, not a deployment story. The same week, The Robot Report ran a state-of-manufacturing package citing the International Federation of Robotics, which found that industrial robot installations doubled in the decade that ended in 2024. The IFR's robot density leaders are South Korea, Singapore, Germany and Japan. Humanoids, the report notes, are just beginning commercial trials while other form factors are already available. So the 77.9% figure needs reading with care. It measures who is shipping units, not who has robots doing paid work at scale.

Two kinds of humanoid factory

The phrase "humanoid robot factory" covers two different things in this week's coverage. One is the factory that builds humanoids, which is where China's shipment numbers come from. The other is the factory that humanoids are meant to work in, which is where the money is supposed to come from eventually. The dossier's sources keep those apart, mostly.

The Robot Report's manufacturing survey points to demand drivers on the buyer side: shortages of skilled labor across electronics, metals, food and beverage and pharmaceuticals. It cites the Association for Advancing Automation, which recorded robot order value up 6.6% year over year in the first half of 2026, with non-automotive orders offsetting a slowdown among automotive OEMs. That is the market humanoid vendors are pitching into, and cobots, autonomous mobile robots and conventional arms serve it today. The trade outlet does not claim humanoids are ready to take that work today.

"While interest in humanoids is high, they are just beginning commercial trials, and other form factors are already available," The Robot Report wrote in its 29 September state-of-manufacturing piece.

That sentence does a lot of work. It is the closest thing in the dossier to a sceptical read on the shipment surge, and it comes from a robotics trade outlet rather than a general news desk.

The home is the harder test

Away from factories, one of the more concrete humanoid launches in the dossier is a domestic robot. Flourish, a startup split between San Francisco and Paris, opened sales for Flourish One, a $3,555 wheeled humanoid-style home robot, according to The Robot Report on 29 September. The first batch is 50 units, hand-built by a small team, and the company says it hopes to ship before Christmas.

Founder Antoine Marcel told The Robot Report that the robot can be taught new chores in under 30 minutes using only a smartphone. The phone's inertial measurement units capture the motion, and the arms mirror it. Skills are fine-tuned per task, per home, with training and higher-level inference running on cloud GPUs. The robot runs on a Raspberry Pi to keep the bill of materials down.

Marcel is direct about why not an NVIDIA GPU on board. "If we were to put an NVIDIA GPU in the robot, it would immediately double the cost," he said. He also explained the choice of the home over industry: "The way I tidy my apartment is not the same as you do in your house. You're going to show it for 30 minutes how to water your plants."

The specification is modest and the company does not hide it. Each arm has a 1.5 kg payload, the battery is designed for 12 hours of arm operation, and there is a small lidar near the floor plus wrist-mounted cameras on the two-fingered grippers. The robot is not waterproof, will not do dishes, and is suited to tidying, trash and wiping a table rather than intricate assembly. Flourish was founded in January of this year.

Where the software is jagged

If the hardware is shipping in volume and the home robots are small-batch, the software layer is the part nobody has clean numbers for. A technical report published on 29 September by fig.inc evaluated frontier models on web automation and four physical domains, including tabletop manipulation and assembly. The authors did not set out to rank vendors, only to test how models behave when the same tasks are repeated.

Its finding is that benchmark averages hide a lot. The authors write that "in every model pair, the lower-scoring model solves tasks the higher-scoring one fails," and that a model update can leave the average almost unchanged while many individual tasks flip. They ran seven models on 177 closed-loop VisualWebArena tasks and three models across Bench2Drive, VLABench, IndEgo and Assembly101, and released item-level results under the name RIDGE.

For anyone reading shipment figures as a proxy for capability, that is the counterweight. A robot can leave the line and still fail a task a weaker model passes, depending on the task category. The report's authors put the noise band for their web benchmark at plus or minus 3 percentage points.

Sensing and the missing sense of touch

Then there is the hardware nobody photographs. ForceN, which supplies force torque sensors for robotic wrists and other joints, will run a session called "A Crash Course in Force & Torque Sensing for Humanoids" at RoboBusiness 2026 on 20 and 21 October in Santa Clara, California, The Robot Report said on 29 September. Founder and CEO Robert Brooks speaks at 2:15 p.m. PT on the first day.

The premise, as the outlet frames it, is that force torque sensors give robots a sense of touch and that without them robots are blind to how much pressure they are applying. ForceN's background is in surgical, logistics, space, nuclear and food and agricultural robotics, and the company says it has supplied sensing systems to humanoid makers over the past three years.

That matters for the factory question more than it first appears. Task categories in the fig.inc evaluation include manipulation and assembly, precisely the work where a gripper that cannot feel will break parts or drop them.

None of the dossier's sources put a figure on how many of the 19,000-plus Chinese units shipped in the first half are doing productive work rather than trials, demos or research. IDC's 2030 forecast of more than 750,000 units is a projection, and projections in this category have moved before. The one number that is not a projection is the 77.9% share, and it says more about where humanoids are assembled than about where they are useful.