Humanoid robots: China's 19,000-unit lead and the factory math behind it

Those IDC numbers landed on 29 September via TechNode, which credited the Chinese outlet IT Home. If the figure holds, China is not merely a participant in the humanoid market. It is the market. The same IDC forecast put global humanoid shipments in 2030 at more than 750,000 units, a number that assumes the current curve keeps bending upward for four more years.

Yet the word factory is doing a lot of work here. Shipping a humanoid out of a building is not the same as putting it to work on a line, and the newest sources in this dossier show how far apart those two things still are.

The manufacturing floor is still mostly non-humanoid

The Robot Report's manufacturing survey, published on 29 September, makes the size of the gap plain. Industrial robot installations doubled in the decade that ended in 2024, according to the International Federation of Robotics. South Korea, Singapore, Germany and Japan lead on robot density, the number of robots per 10,000 workers. In the first half of 2026, robot order value rose 6.6% year over year, according to the Association for Advancing Automation. The group noted that growth in non-automotive sectors offset a slowdown among automotive OEMs.

Humanoids appear in that survey as a promise, not a line item. The report states that while interest in humanoids is high, they are just beginning commercial trials, and other form factors are already available. Cobots, autonomous mobile robots and mobile manipulators are the shapes actually doing the work today.

So the headline number from IDC and the on-the-ground picture from the trade press are not in conflict. They are measuring different things: units leaving a factory versus units doing useful labour inside one.

Force and torque sensing is the unglamorous bottleneck

If the Chinese shipment figure describes supply, a talk scheduled for RoboBusiness 2026 describes the demand-side problem. ForceN founder and CEO Robert Brooks will present "A Crash Course in Force & Torque Sensing for Humanoids" at 2:15 p.m. PT on 20 October in Santa Clara, California, the first day of the two-day event, according to The Robot Report.

Force torque sensors give robots a sense of touch. Without them, as the same report puts it, robots are blind to how much pressure they are applying while working. Brooks is expected to draw on three years of work with humanoid companies and on force and torque development in surgical, logistics and field robotics. His stated aim is to accelerate production and field readiness.

That is a useful corrective to shipment-count coverage. A robot that can grip a box is not a robot that can grip an egg, and the difference lives in the sensing stack, not in the unit tally.

A $3,555 home robot, and what its spec sheet admits

The domestic side of the market is moving too. Flourish, a startup split between San Francisco and Paris, opened sales for Flourish One, a wheeled humanoid-style home robot priced at $3,555, The Robot Report reported on 29 September. The company says early adopters can teach it new chores in under 30 minutes using only a smartphone.

The specification is unusually candid about limits. Each arm has a payload capacity of 1.5 kg. The battery is designed for 12 hours of arm operation. A small lidar sits on the base near the floor, and each two-fingered gripper carries a wrist-mounted camera. The robot is not waterproof, so it will not do dishes or wash a dog in this generation. It is suited to tidying, trash and basic cleaning such as wiping a table rather than intricate assembly.

Flourish is building the first batch of 50 units by hand and hopes to ship before Christmas. To keep the bill of materials down, the unit runs on a Raspberry Pi, with cloud GPUs handling training of skill models and inference for higher-level behaviour. A network connection is required for "thinking," or users can optionally run workloads on their own computers. Founder Antoine Marcel told The Robot Report that putting an NVIDIA GPU in the robot would immediately double the cost, and that 1.5 kg was chosen because most home tasks do not need a huge payload.

"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," Marcel said. "We're going to fine-tune an AI model for that... now the robot is able to do that in your house."

The teaching method uses the inertial measurement units inside a phone. The user holds the phone and moves it as if performing the task while the robot mirrors the action. Marcel called that the magical part of the product, because no extra hardware is needed.

The software is jagged, and the benchmarks hide it

Hardware limits are only half the story. A technical report published on 29 September by researchers at fig.inc evaluated frontier models on web automation and on physical tasks including manipulation, assembly, industrial procedures and driving. The authors call the results jagged performance.

The finding is blunt: in every model pair tested, the lower-scoring model solved tasks the higher-scoring one failed. On web tasks, the best model changed with the website and with the benchmark's own difficulty labels. A model update could leave the average almost unchanged while many individual tasks flipped. An update that raised the average still lost tasks the older model had solved.

The same unevenness held in physical domains. Models that beat Gemini 3.5-Flash on every domain average still did worse on some task categories. The authors released a dataset called RIDGE with item-level results across five domains, from VisualWebArena to Bench2Drive, VLABench, IndEgo and Assembly101. Their point is that a benchmark mean hides how much success varies task by task, which matters when deployment decisions rest on those numbers.

For anyone reading shipment totals as a proxy for capability, that is the caveat to carry. Volume tells you how many units left the building. It does not tell you how reliably any of them will water a plant.

What the next twelve months will test

The dossier's newest sources point in one direction: the humanoid conversation is shifting from demonstrations to specifications. Payload in kilograms. Hours of operation. Whether the robot needs a network connection to think. Whether a sensor package can keep it from crushing what it holds.

China's 77.9% share of first-half shipments, as reported by TechNode, is a real and striking data point. IDC's 750,000-unit forecast for 2030 is a projection, not a shipment. Between those two numbers sits the harder question that The Robot Report's manufacturing survey raises: how many of these machines will be doing paid work on a factory floor or in a home, rather than counting toward an installation statistic.

RoboBusiness 2026, on 20 and 21 October in Santa Clara, will put some of that on the record. ForceN's session is one data point. The rest will show up as order values, field trials and, eventually, whether a $3,555 robot can be taught to water the plants without being taught twice.