BLOGS

Peak torque sells actuators. Duty cycle buys robots. Through October 2026 the humanoid supply chain has been arguing, loudly, about which transmission wins the joint — harmonic, planetary, cycloidal, or no gearbox at all. The more useful engineering conversation is a different one: whichever mechanism you pick, the joint will be sized by the continuous torque it can hold across an entire motion cycle, and by how much heat it can shed while doing it. This article works through that sizing problem the way a program team actually faces it — from the duty cycle, to the joint map, to the feedback stack — and explains why integration became the dominant architecture of 2026.

The Number That Is Not on the Datasheet

Walk into any actuator selection meeting and the first figure on the table is peak torque. It is also, usually, the least important one. Peak torque is a magnetic and thermal limit that can be touched for a fraction of a second; a humanoid in commercial service will run thousands of cycles a day and will spend most of each cycle in the low-speed, high-torque region where copper loss is highest and cooling is weakest.

The specification that decides whether the joint survives is continuous RMS torque: the equivalent constant torque that produces the same heating over the whole motion profile. A sound workflow is now well established among integrators and looks roughly like this:

• Compute the RMS torque from the complete motion cycle — not the showcase manoeuvre, but the repetitive one.
• Add a dynamic margin for impact events (1.5× is a common engineering convention for gait landing and sudden stops).
• Check the reducer's permitted peak torque and the drive's overload curve at that duty point.
• Then check the winding temperature rise, because that is the constraint that binds first in practice.

A joint that must lift 8 kg at 0.9 m of reach needs about 72 N·m of static torque before any dynamics are considered. But a factory humanoid does not lift once; it lifts, holds, and lowers against gravity for an entire shift, and never fully returns to ambient between cycles. In that regime the limiting factor is rarely the torque rating printed on the module — it is how fast the winding can get rid of the heat that holding torque generates. This is the practical reason the industry's centre of gravity has shifted from "how much torque per kilogram" toward "how much torque per kilogram, sustained."

Actuators Are the Bill of Materials

The intensity of this debate makes sense once you look at where a humanoid's cost and mass actually sit. Industry trackers consistently place the actuator at the single largest cost block in a humanoid bill of materials, commonly quoted in the range of 40–60% depending on the joint count and transmission mix. Chinese supply-chain analysis published in early October 2026 narrows the same point to the integrated joint module alone, putting it at roughly 40–50% of total hardware cost — which is why "mass production and cost reduction" has become the explicit main line of every domestic joint program.

Mass is the other half of the story. One widely cited estimate attributes around 56% of a humanoid's total weight to its actuation system (IDTechEx, cited in industry selection guides). With a production humanoid typically carrying somewhere between 25 and 40 actuated joints, every gram saved in a reducer housing or motor stator propagates through the whole kinematic chain: lighter distal segments mean lower reflected inertia at the hip, which means less torque demanded from the largest actuators, which means a smaller cell and a lighter battery. Actuator mass is not a local optimisation; it is a whole-body one.

The Joint Map: Different Joints, Different Machines

The single most damaging habit in actuator selection is treating "the humanoid joint" as one product. It is not. A wrist and a knee occupy opposite corners of the design space, and the transmission that wins one loses the other.

The bands above reflect the selection frameworks used by integrated-joint suppliers and are best read as engineering conventions rather than standards. What matters is the pattern: the distal joints trade torque for compactness, and the load-bearing joints trade compactness for shock tolerance. Reducer market analysis points the same way — the core ratio band for humanoid joints sits around 50:1 to 100:1, and strain-wave (harmonic) gearing is the largest single type by volume, precisely because it fits the numerous, low-torque, high-precision joints that make up most of a robot's degrees of freedom. Our own engineering breakdown of harmonic, planetary and cycloidal drives across the joint map covers that segmentation in more detail.

Why Integration Is a Thermal Strategy

The integrated joint module — frameless torque motor, reducer, encoder, brake, drive electronics and increasingly a torque sensor in one sealed housing — is sometimes described as a packaging innovation. It is more accurate to call it a thermal and electrical one.

Two effects matter. First, integration removes interfaces. Each cable, connector and separately mounted drive is a path for EMI, a source of tuning drift, and an assembly step that has to be qualified. A module with one control connector per axis instead of five is not just faster to build; it is measurably more repeatable, which is what makes automated calibration possible at all.

Second, and less obviously, integration lets the thermal design be solved once, globally. In high-torque-density frameless motors (12 to 14 pole pairs is typical), the winding-to-housing path, the housing-to-air path and the duty cycle are co-designed rather than bolted together. That is why the thermal question and the integration question are the same question: you cannot optimise a duty cycle you do not control.

The market has already voted. One 2026 market outlook estimates that integrated servo joints account for roughly 78% of shipments, with mainstream torque density now above 35 N·m/kg — about 20% higher than in 2024. Those figures come from a single research house and should be treated as directional, but the direction is unambiguous: the discrete motor-plus-gearbox-plus-external-drive architecture is losing share to the sealed module.

The Quasi-Direct Drive Trade-Off

No discussion of joint architecture in 2026 is complete without quasi-direct drive (QDD) and its cycloidal variant. The appeal is mechanical and control-theoretic at once. Use a low reduction ratio — usually under 10:1 — and the motor current stays nearly proportional to output torque, so the joint can estimate force from current without an expensive torque sensor. Reflected inertia scales with the square of the ratio, so a low ratio keeps the leg light and responsive. And when a foot slams down, a low ratio lets the impact back-drive the rotor instead of shearing teeth inside a reducer.

The trade-off is equally clear, and it is a duty-cycle trade-off. Low ratio means the motor itself must produce the joint's torque, which raises the demands on magnet grade, winding design and cooling, and asks the drive for high peak current. And there is a specific failure mode that only appears in sustained operation: holding a pose. A knee that must keep a humanoid standing still has to output continuous torque at near-zero speed — the worst possible thermal operating point. Some teams answer with pulsed torque-holding strategies that let the winding "breathe" between high-current intervals, but the underlying physics does not go away.

The cycloidal variant of QDD (sometimes written C-QDD) tries to buy back some of that margin with a low-ratio cycloidal stage — around 10:1 — paired with a high-torque BLDC motor. Our earlier analysis of where direct drive ends and gearing begins reached a similar conclusion by a different route. Research from UCLA and other groups suggests the cycloidal tooth profile distributes stress better than planetary teeth under impact, which is exactly the regime legged and humanoid robots live in. The cost is manufacturability: cycloidal geometry demands tight tolerances (electro-discharge machining at the ±2 µm level is often quoted) and is prone to torque ripple and vibration from rotor eccentricity. It is an emergent answer, not a settled one.

Where Cycloidal Comes Back

For the load-bearing joints, the pendulum has swung back toward cycloidal gearing — the family that predates the humanoid boom by decades. In September 2026, one veteran Japanese cycloidal-gear maker's first-generation precision reduction gear was registered in Japan as an Essential Historical Material for Science and Technology, a reminder that compact, high-rigidity cycloidal gearing has been solving shock and torque-multiplication problems in industrial robots for a very long time. The same manufacturer states that its precision reduction gears hold roughly 60% of the global market for medium- and large industrial robot joints.

Humanoid robotics is not replacing that technology; it is giving it a new application frontier. Hip and knee joints need what cycloidal stages have always offered: high stiffness, very low backlash, and the ability to absorb transient loads well above rated torque without permanent deformation. The engineering challenge is adaptation, not invention — shrinking the stage, integrating the bearing, and making it compatible with the sensing and communication architecture of a modern joint module.

What the Feedback Stack Buys

If torque and thermal budget define what a joint can do, the feedback architecture defines how well it can be controlled — and, increasingly, how safely. Two encoders are now common in precision joints: a high-resolution encoder on the motor shaft for commutation and velocity control, and a second encoder (or an inductive/resolver sensing element) at the joint output for true output position. The difference between the two readings is torsional deflection in the transmission, which can be used to estimate torque without a dedicated sensor, or to detect faults. We wrote about the accuracy ceiling this dual-encoder architecture sets in an earlier engineering note.

Adding a real torque sensor — a strain-gauge or inductive element measuring the actual joint moment — changes the control regime. The joint can implement impedance control: it responds to contact with a defined stiffness and damping rather than executing a rigid position command. That is the difference between a robot that can shake a hand and one that can only reach a coordinate. It is also what makes a joint safe in a shared workspace, because force limits become programmable rather than structural.

The practical caveat is that every sensing channel adds calibration burden. The reason the integrated module has won is not just that it bundles components; it is that bundling them makes the calibration of that whole stack a factory process rather than a field exercise.

From Prototype to Line

The last year's most consequential shift is not on the datasheet at all. It is that integrated joints moved from hand-assembled samples to automated production, and the numbers from early October 2026 show how fast. One Chinese actuator maker's integrated-joint line reports automation above 85%, a takt time of roughly 90 seconds per module, and first-pass yield consistently above 96%, with the product family spanning a 2–400 N·m torque range. Another supplier reports delivering close to 95,000 joint modules in 2025 against a planned line capacity on the order of 100,000 units a year. These are company-sourced figures and should be read as such — but the mechanism behind them is the real story.

Automating joint assembly forces three things to become explicit: the tolerance stack, the calibration procedure, and the test that proves both. That is why automated lines and consistently low backlash are correlated in practice. A reducer with 5 arcmin of backlash and a calibration procedure that can be repeated identically 100,000 times is worth more to a program team than a hand-built sample with 1 arcmin and no documentation, because only the former can be designed into a robot that ships.

For anyone specifying joints in 2026, the industrialisation question has become part of the technical question. Ask not only what the module can do, but how repeatably the same module can be produced, calibrated and verified — because with 25 to 40 joints on every robot, the variance across a production population is the number that reaches the customer.

Ti5ROBOT builds integrated rotary joints and harmonic joint modules and humanoid robot actuators, and the engineering view above is reflected in how we specify and qualify our own modules — duty-cycle first, integration as a thermal strategy, and calibration as a factory process. If you are sizing joints for a humanoid or collaborative platform, the conversation is more useful when it starts from your motion profile than from a peak-torque number.

Source: Industry selection frameworks and 2026 market outlooks on humanoid rotary actuators and speed reducers; Chinese supply-chain reporting on integrated joint module mass production (October 2026); IDTechEx weight-share estimate as cited in supplier selection guides; cycloidal quasi-direct drive research reported via market and preprint sources. Figures attributed to single sources are labelled as such in the text.