Industrial robots already do the work factories were rebuilt to give them. They take a task that repeats, sit in a space designed around the arm, and earn their keep as infrastructure.

Humanoids are being asked to do something harder. They are sold as general-purpose machines that can walk into a workplace built for people and take the jobs that fixed automation cannot take cheaply enough to justify a rebuild.

This analysis looks at who is buying them, what jobs they are being asked to do, and whether the economics hold.

An industrial robot, here, is the fixed or specialised arm already common in electronics and car plants. A humanoid is a two-armed, mobile machine sold as able to work in rooms designed for people.

A listing can price a maker. It cannot prove the job.

On 19 August 2026, Hangzhou-based Unitree listed on Shanghai’s STAR Market as 688836.SH. TechNode, reporting the next day from the listing record, put 2025 revenue at CNY 1.699 billion. Humanoid products accounted for CNY 868 million, more than half the company. Profit attributable to shareholders was CNY 278 million. The offer priced at CNY 150.80. Proceeds were about CNY 6.099 billion.

Two years earlier, humanoid sales were almost a rounding error. The same record puts 2023 humanoid revenue at CNY 2.96 million, less than 2 percent of the firm. The prospectus, as restated by the Global Times, also claims more than 5,500 humanoids shipped in 2025.

That makes humanoids a serious industrial category in commercial terms. Public markets are now willing to value a humanoid maker as a company, not a prototype lab. What the listing does not show is whether humanoids have become productive infrastructure. A public valuation can put a price on the category but it cannot prove that a general-purpose body has taken a factory job.

For that, you need a buyer.

The first commercial job is inserting cushioning

On 3 September 2026, CJ Logistics said it had put two dual-arm humanoids on a live packaging line at its Yangji Olive Young centre in Yongin, Gyeonggi. The job is narrow: machines inserting cushioning. A worker still packs the ordered goods. The body is from Robotis, the hand from Aidin Robotics, the foundation model from RealWorld AI. CJ calls this a first in Korean logistics.

Yangji is a live line where a person still finishes the box. Olive Young is a cosmetics retailer, and its fulfilment line handles many small items and changing product mixes. That makes it difficult to automate with a machine designed around one fixed task. A humanoid is being tested because it can work in the same space as a person.

Kim Jeong-hee, at CJ’s TES Logistics Technology Research Institute, described the move as a step out of demonstration and into live operations. Herald Business put the operational reason more plainly: fixed automation struggles when the product mix and the process change. A humanoid is supposed to use the workspace already built for people. In short: do not rebuild the room. Send in a machine that can work in the room you have.

The numbers that matter still come from Gunpo

Yangji is the live line. The economics CJ has been willing to discuss sit earlier, at Gunpo, in the second half of 2025.

In a 10 June 2026 interview with the Korea JoongAng Daily, Koo Seong-yong, head of automation technology at CJ Logistics, put the Gunpo pilot at 30 to 50 percent of an experienced worker’s speed, or about 130 boxes an hour. A dropped piece of cushioning still had to be picked up by a person. A unit cost about KRW 150 million, including development, against about KRW 50 million a year in site labour. Break-even would take about three years if the machines got faster. Continuous running, minus charging time, was the offset for slower unit speed. The target for the second half of 2026 was about 70 percent of human pace and 99 percent accuracy.

Those figures describe a 2025 field test. Yangji has not published its own boxes-per-hour. After the pilot, CJ bought two more units, with three planned. Koo also put fulfilment-centre automation at about 30 percent: palletising arms and autonomous mobile robots on the standardised work. Packaging and mixed-SKU sorting remain mostly human. That is why the first commercial job is stuffing foam, not running a caged assembly cell. The obvious work already has a machine. The leftover work is what a humanoid is being asked to try.

On 8 April 2026, the International Federation of Robotics put Korea at 1,220 industrial robots per 10,000 manufacturing employees, the highest figure in that release. Electronics and cars drive it. Singapore, in the same note, stands at 818; Japan at 446. That density measures fixed automation where the task is predictable. It is not a warehouse-humanoid census.

Against that installed base, two machines stuffing foam at Yangji is a small operation. It is also the right kind of small. The country that already automated the repeatable factory jobs is now testing a different machine on the work those jobs never absorbed.

The room is the difference

An industrial robot replaces a task by redesigning the environment around the machine: fixtures, cages, programmed paths. The productivity is real. So is the rebuild. A humanoid is sold on the other proposition: take the task without asking the workplace to be rebuilt around it.

That is why the leftover work can be commercially interesting in a country that already has 1,220 industrial robots per 10,000 manufacturing employees. The arms took the predictable jobs. What remains is too variable to justify a fixed cell, and still structured enough for a machine to learn.

The constraint is capital against speed. A buyer will keep a KRW 150 million machine only if it closes enough of the gap with a KRW 50 million worker. The 70 percent target and the three-year payback sit at the centre of that test. They are Gunpo numbers. Yangji has not published its own. The June math only works if a machine running at 30 to 50 percent of human speed gets closer to 70 percent.

A humanoid that keeps the room intact can still be the expensive machine if it stays slow. Category scale in China and a live line in Korea do not collapse that difference. They make it easier to see.

A listing in Hangzhou can make humanoids look like an industry. Two machines in Yongin make them look like an operation. Korea’s industrial-robot density is why the operation is interesting: the leftover work is the market.

What to watch

The first test is still at CJ. Does performance on the live line move toward 70 percent of human pace, and does the capital case still look like a three-year payback once the machines are no longer a pilot? A published Yangji throughput number would be the first clean read. Until then, the 70 percent target remains a 2026 aim from June, not a result.

The second is whether other buyers show up with named sites, repeat orders and identifiable production tasks. That would move the category from a capital-market story, built around listings, product cycles and shipment claims, to an industrial market.

The third is whether humanoid work expands into tasks that conventional automation has not already captured. More humanoids on the same cushioning job would still be a narrow proof. Humanoids on work that a fixed cell cannot take cheaply would be a different proof.