Humanoid Robot Factory: Five Million Industrial Robots, Only Thousands Are Humanoids

Two numbers, published four days apart, describe the gap between the humanoid robot factory and the factory robot. The IFR put the operational stock of industrial robots at 5 million units in 2025, up 9%, after factories installed more than 600,000 new machines last year. IDC counted fewer than 25,000 humanoid units shipped globally in the first six months of 2026.

That is not a rounding error. It is the entire humanoid industry, including China, which accounted for 77.9% of those shipments and more than 19,000 units on its own, per IDC figures republished by TechNode on 29 September.

What the factory floor already looks like

Before humanoids get a shift, the boring form factors are doing the work. Maven Robotics, founded in 2024 in Santa Clara, has built four iterations of a wheeled mobile manipulator, raised $100 million, and closed a production contract with its first customer. Its robots run 16 hours a day, five days a week, with up to 99% uptime, according to Hamza Derbas, the company's co-founder and CEO, speaking to The Robot Report on 1 October.

Maven's pitch rests on a simple objection to legs. "At the same time, we noticed that the humanoid robotics companies, the ones with legs, were starting with the human form factor and trying to work backwards into a problem to be solved," Derbas told The Robot Report. The company says its base fits through doors and reaches 3 m (9.8 ft) into shelving while carrying 30 kg payloads. It also runs a shared-learning network: "If a robot makes a mistake somewhere in California, a robot in Japan can learn how to deal with that mistake the next day," Derbas said.

The Chinese numbers point the same way. China installed 354,000 industrial robots in 2025, up 20% year on year and 59% of global deployments, according to the IFR release dated 24 September. Domestic suppliers took a 55% share of that home market, up from 57% a year earlier, no, down: the IFR records 55% in 2025 against 57% in 2024. Japan fell 19% to 36,219 units and lost its second-place ranking to the United States, which installed almost 38,500 units, a 12% rise.

The humanoid assembly line is the hard part

Tesla has the most detailed public account of what goes wrong when a carmaker tries to build people. Ars Technica reported on 26 September that Tesla's Fremont factory stopped making the Model S and Model X in May 2026 and moved line workers and engineers onto Optimus, citing extensive reporting by The Information. The same account describes Optimus V3 production running at hundreds of robots per week against a target above 1,000 per week by the end of 2026.

The bottlenecks are small. Optimus hands and forearms together contain more than 100 components such as screws, assembled by hand, and Ars Technica reports that newly produced robots need immediate fixes. Touch sensors for the hands proved unreliable enough that Tesla developed a glove-like sensor layer that can be swapped without replacing the whole hand. The Information's sources described the robots as requiring programming for specific tasks in controlled environments.

Ars Technica also reported that Tesla asked factory workers in Texas and California to wear suits that record their physical movements, an imitation-learning play, and that some workers pushed back against training their replacements. Musk described Optimus during Tesla's second-quarter 2026 earnings call as potentially "the biggest product ever," while conceding that an autonomous humanoid capable of many tasks is "one of the hardest things to solve."

Nobody has solved the hand. ForceN founder and CEO Robert Brooks will give a talk titled "A Crash Course in Force & Torque Sensing for Humanoids" at RoboBusiness 2026 on 20 and 21 October in Santa Clara, drawing on three years of work with humanoid companies, The Robot Report noted on 29 September. Force and torque sensing is what lets a robot know how hard it is gripping. Without it, the machine is blind to pressure.

Vishay Precision Group makes the same argument from the component side. The company will present custom force and torque sensors at RoboBusiness, building strain gages directly into structural parts, The Robot Report reported on 1 October. Ron Zukerman, VPG's innovation director, is quoted saying the company has "sixty years of experience in aerospace-grade force sensing." Off-the-shelf sensors, the company argues, are too large for small limbs, have measurement ranges that are too broad, and need bulky external electronics.

Data, memory and the unglamorous stack

Training data is the other wall. Innodata opened a motion-capture lab on 30 September to generate 3D training data for humanoids and industrial robots, The Robot Report reported. CEO Rahul Singhal said physical AI teams hit a common ceiling: "There isn't enough real-world interaction data, and what exists is expensive and slow to produce." Franklin Tanner, the company's vice president of robotics and physical AI, told the outlet that a humanoid weighing almost 200 lb (90.7 kg) needs readings that are precise, not in the ballpark.

Memory is a constraint too. A sponsored article in Semiconductor Engineering on 1 October cites Goldman Sachs analyst reports projecting annual humanoid shipments above 6 million units by 2035, up from fewer than 15,000 in 2025, and notes that logging motor torque, velocity and limb position above 10 kHz sampling rates can exceed the endurance limits of conventional non-volatile memory. Robots need distributed flash with power-loss resilience, not just a big AI chip.

Not every humanoid is chasing the factory. Flourish, split between San Francisco and Paris, opened sales for Flourish One, a $3,555 wheeled home robot with 1.5 kg payload per arm and a Raspberry Pi inside, The Robot Report reported on 29 September. Founder Antoine Marcel told the outlet he is betting homes will fine-tune skills task by task rather than waiting for a generalist model. The first batch is 50 units.

What the numbers do not settle

Two forecasts disagree about where this goes. IDC raised its projection for global humanoid shipments in 2030 to more than 750,000 units, according to TechNode's 29 September summary. Goldman Sachs, as cited in the Semiconductor Engineering piece, puts the 2035 figure above 6 million. Neither number is a production schedule.

What the IFR data does show is that the factory floor is already automated, and mostly not by humanoids. Five million industrial robots were operating worldwide at the end of 2025, the federation said, with China holding 59% of new installations. Humanoids are a rounding error against that base, and the engineering problems that keep them there, hands, sensors, training data, memory endurance, are being worked on in public by companies that have not yet shipped at scale.

Bruno Siciliano, a robotics professor at the University of Naples Federico II and a 2024 IEEE Robotics and Automation Society Pioneer Award winner, put the case against the fully anthropomorphic factory robot in Il Sole 24 Ore on 1 October, pointing to BMW's rethink of its use of Figure AI robots in South Carolina. A fixed base with two arms, or a robot on wheels, may simply be more productive. Siciliano also argued for a European physical AI model built around human-robot interaction rather than replacement, and said a proposed AI-based robotics act with a 30 billion euro budget should target a 33% share of the world market by 2035.

On the same day, Figure showed what happens when a generation ends. The US robotics company retired its F.02 fleet in favour of F.03, sending units into a 75-tonne electric arc furnace in Imatra, Finland, with Arnold Schwarzenegger involved and former MythBusters members helping plan the stunt, heise online reported on 1 October. The company said keeping the old robots operational was too costly and dismantling them too slow, and that protected components had to be destroyed. How many units went into the furnace is unclear.