A humanoid robot is usually described by its joints, and on that measure China's chain is moving. Kinco shipped 113,600 frameless torque motors in the first half of 2026, and Leaderdrive sold 425,158 harmonic reducers in 2025. The hand tells a different story. It is where the robot has to grip, feel force and handle objects without breaking them, and the filings show a stage that is still sampling, not shipping at scale.
What a hand needs
A hand is not one part but many small ones. Zhaowei describes the two main designs: fully driven hands, with a micro drive module in each finger joint, and under-actuated hands, which use linkages to grip with fewer motors (2026 interim report, p. 10). The robot also needs to feel force, at the wrist and at the joints, to handle objects safely. That is the job of six-axis force and torque sensors and joint torque sensors, the line Keli Sensing sells (2026 interim report, p. 8).
So the hand stage has two halves, and in the first half of 2026 only one of them reported any units at all.
Sensors: thousands, and a long sample list
Keli is the one company on this map that puts a number on robot sensing. It says its robot force sensor business moved from small batches to batch supply in the first half of 2026, that monthly orders passed 1,000 units from April, and that it sold more than 2,000 robot force sensors in January to June, more than its shipments in all of 2025 (p. 9). It has sent samples to more than 90 humanoid, collaborative and industrial robot makers, some have placed batch orders, and several leading customers are validating its sensors. Batch shipments are going to AI massage robots and humanoids.
What it does not give is just as telling. It reports no robot revenue line, so the sensor business cannot be sized against its ¥837.8m of half-year revenue, and it names no humanoid customer. It has invested in companies in the robot chain, among them the humanoid maker Kepler, but an investment is not a supply contract, and the report does not say Kepler buys its sensors.
Hands: products on sale, no numbers attached
Zhaowei sells hands on both technical routes, labelled A17, B21, B06 and C06, and sells the drive modules separately. It says commercialisation of hands and core drive modules is advancing and that it holds a strong position in the segment (pp. 10, 11). The report gives no hand or module revenue, units or customers. Its own results are weakening, with revenue up 3.8% to ¥816.4m and net profit down 34.1% to ¥74.6m in the half, and the company does not tie that fall to robots (p. 7).
Leadshine lists dexterous hands with high, medium and low degrees of freedom among its robot parts, and says its frameless torque motors, high-degree-of-freedom dexterous hands and joint modules made market progress in the half, with order size and customer coverage up "significantly" (2026 interim report, pp. 9, 11). It gives no robot revenue, units or customers either, so its robot business cannot be sized from the report. Its growth in the half, revenue up 39.9% to ¥1.25bn, came mainly from factory automation.
Haozhi lists torque sensors and encoders among its robot parts and Inovance lists sensors and encoders among its humanoid parts, but neither gives revenue for them. Across the whole stage, then, the public record is one sensor maker shipping in the thousands, two hand makers selling without figures, and no named humanoid customer anywhere.
Why the gap matters more than the joint gap
The joint stage has the same disclosure problem in places: no filing names a humanoid customer for frameless motors, and the car-parts makers in our earlier article report robot revenue that is a rounding error. But joints at least have volumes from industrial customers to fall back on. Kinco's 113,600 motors and Leaderdrive's 425,158 reducers are real shipments today, even if not all of them go into humanoids (2026 interim report, p. 15).
Hands and sensors have no equivalent fallback disclosed. A force sensor sold to a massage robot or a collaborative arm is a sale, but it does not prove the part works in a humanoid hand, and a hand with no reported units has not yet shown it can be made at a price a robot maker will pay. The hand is also the part a buyer sees: a robot that walks well but cannot pick things up reliably is a demonstration, not a worker. That makes the hand stage the likeliest place for a robot maker to keep design in-house, as Unitree already does with its joint motors and drives, rather than buy a complete hand from a supplier that cannot yet show volumes.
Our view
The hand stage is earlier than the rest of China's humanoid chain, and the evidence points the same way from every company that touches it. Sensors are furthest along, with batch orders and a monthly run rate above 1,000 units, but still unnamed customers and no revenue line. Hands are a step behind that: products exist on both technical routes, but no maker will say how many it sold, for how much, or to whom. Until one of them does, the hand is a claim about capability, not a business, and it is the stage most likely to decide whether a humanoid is useful rather than merely impressive.
What to watch
- Whether Keli's monthly sensor orders hold above 1,000 units in the second half of 2026, and whether it names a humanoid customer.
- A reported hand or module revenue line from Zhaowei or Leadshine, which would size the stage for the first time.
- Whether a robot maker names an outside hand supplier, or whether hands follow joints in-house.
- Keli's sample list: more than 90 makers sampled, but only a handful in validation, so conversion, not sampling, is the number that matters.
The full picture is on our humanoid robot industry map, with each stage's companies, volumes and margins.