BLOGS

Axial-flux "pancake" motors are entering robot joints on a promise of two to four times the power density. The honest version is narrower and more useful: they buy torque per millimetre of axial length, not torque per kilogram — and the invoice arrives as rotor inertia, a two-kilonewton magnetic pull, and losses that climb with electrical frequency. Here is the arithmetic behind the decision.

The claim, stripped of adjectives

At the 2026 China International Industry Fair, flat-disc joint motors were everywhere: hip modules, knee modules, and a growing shelf of axial-flux units marketed specifically at humanoid makers. The accompanying numbers have a familiar shape — half the volume and weight for the same output, two to four times the power density of a conventional radial machine, torque density up by more than 120 percent.

Those figures are not fabricated. They are also not comparable. A "two to four times" claim without a named baseline motor, a stated duty cycle and a stated cooling condition is marketing arithmetic, not engineering data. The same trap sits inside specific-torque comparisons: a joint-level figure of 100 to 200 Nm/kg describes an assembly with a reducer attached, while a bare-motor figure of 2 to 5 Nm/kg describes magnets and copper with nothing bolted on. Both are true, and subtracting one from the other tells you nothing.

So start with the one comparison that is like-for-like. A 2026 IEEE/ASME Transactions on Mechatronics study optimized an axial-flux motor paired with a cycloidal reducer for wearable robotics, and published a table of competing designs. Its own actuator reaches 45 Nm peak from roughly 600 grams in a 33 mm thick package — about 75 Nm/kg. Two radial-flux entries in that same table reach roughly 90 and 120 Nm/kg. On specific torque, the axial design loses.

What it wins is thickness: 33 mm against 63 mm and 75 mm for the radial entries. That is the entire defensible claim, and it is a good one — just not the one on the banner. It also explains why the same debate keeps resurfacing in different clothes: as four levers of power density, where torque per kilogram and torque per litre pull in opposite directions.

Two scaling laws that explain the rest

The difference is geometric, and it falls out of the torque integral. Torque in any electric machine is shear stress integrated over the air-gap area, multiplied by the radius at which it acts:

T = ∫ σ · 2πr² dr = (2π/3) · σ · R³ · (1 − k³)

where k is the ratio of inner to outer radius. At fixed thickness and fixed shear stress — meaning fixed cooling capability — torque scales with the cube of diameter. A radial machine integrates the same stress over a cylindrical surface instead, so its torque scales with D² times stack length.

The practical consequence is a design lever you can compute before you cut metal. Moving a joint motor's outer diameter from 80 mm to 100 mm raises axial-flux torque by (100/80)³, or +95 percent, with no change in thickness. The same diameter change in a radial machine yields (100/80)², or +56 percent, and you still need more stack length to close the gap. Wide bore beats long stack — that is the physics behind every flat joint motor on a trade-show floor.

The real currency is axial millimetres

Humanoid joints are not usually mass-constrained. They are volume-constrained, and specifically axial-length constrained. A knee has to fit inside a leg whose silhouette is set by the shin, the thigh shell and the ankle. Inside that envelope, the axial stack has to accommodate the motor, the reducer's flexspline and wave generator, two encoders, the driver board, a brake, and the cable corridor.

Thirty millimetres of saved axial length is therefore not a cosmetic win. It is the budget that pays for a second encoder on the output side, or a wider hollow bore for harness routing, or a larger flexspline for a given torque, or a lower reduction ratio. That last option is where axial flux becomes strategically interesting rather than merely compact.

A lower ratio improves backdrivability, efficiency and shock tolerance, but it demands more motor torque. Cubing diameter is the cheapest way to buy that torque without adding length — which is why flat motors and low-ratio joints keep appearing together in the same designs.

Cost one: rotor inertia, multiplied by N²

A thin disc rotor has moment of inertia J = m·D²/8 about its axis. Real rotors with magnets and hubs run 10 to 30 percent above that, but the scaling is what matters: at constant mass, inertia grows with the square of diameter. Putting the same mass at a larger radius is expensive.

Now add the transmission. Motor inertia reflects to the joint output multiplied by the square of the reduction ratio: J_reflected = J_motor × N². This is the term that decides whether a joint can be force-controlled by reading motor current, or whether it needs an output torque sensor to know what it is doing.

Read the last column slowly. A large-diameter rotor behind a 100:1 harmonic drive contributes fifteen kilogram-metres-squared to the joint's apparent inertia — one to two orders of magnitude above the inertia of the leg link it is supposed to move. At that point the actuator no longer feels like a muscle attached to a limb; it feels like a flywheel with a limb bolted on. Collision detection, impedance control and any control law that assumes the load is the dominant inertia all degrade, and the fix is either an output-side torque sensor or a much lower ratio.

This is the single most common misapplication of axial flux. A wide disc behind a high ratio multiplies the exact penalty the wide disc created. The topology is at its best behind a 6:1 to 15:1 stage, where N² stays in double digits and the diameter penalty is paid once rather than squared.

Cost two: the two kilonewtons trying to close the gap

A flat air gap is a magnetic bearing that nobody designed. The attractive pressure across it is p = B²/(2μ₀). Take a 0.8 T gap over a 100 mm disc — 7.85 × 10⁻³ m² of area — and the pull comes out at roughly 2,000 newtons per gap face. That is about 200 kilograms of force, permanently squeezing the rotor toward the stator, on a machine whose entire performance depends on holding a sub-millimetre gap flat.

Two consequences follow. First, the bearing stack has to carry that preload plus the joint's external loads plus shock, without deflecting the rotor enough to rub. Second, housing stiffness stops being a weight-optimization exercise and becomes a hard constraint: a cover that flexes 50 μm under landing impact is a cover that closes a 250 μm gap by a fifth.

The standard answer is a double-gap topology — one stator between two rotors, or one rotor between two stators — where the two attraction forces are nearly equal and opposite, so the bearing sees only the imbalance and eccentricity components. The price is that the number of critical gap faces doubles, parallelism and concentricity tolerances tighten, and the assembly process window narrows. For a production line, that is often the real cost of the architecture, not the magnets.

Write axial stiffness in N/μm and gap tolerance in μm into the joint specification, next to the torque numbers. A pancake motor datasheet that lists only torque, mass and diameter has told you nothing about whether it will survive a knee landing.

Cost three: losses migrate to frequency

Flat, high-pole-count machines run at high electrical frequency. Core loss follows the Steinmetz form P = k · f^α · B^β, with α between roughly 1.5 and 2 — hysteresis rising linearly with frequency, eddy currents with its square. A 40-pole machine spinning at a few thousand rpm reaches electrical frequencies where a thick laminated core turns into a heater, and where the winding's AC resistance climbs well above its DC value.

Three mitigations, each with a bill attached:

• Segmented, thin-lamination stators cut eddy loss but complicate winding and raise cost per unit of torque.
• Soft magnetic composites suit the three-dimensional flux path of an axial machine but trade permeability and loss against moldability.
• Ironless PCB stators eliminate core loss entirely. A 2025 arXiv study of a 19 mm PCB axial-flux motor achieved 158 mN·m of stall torque with a 0.25 mm gap and a 61.7 percent gain in stall torque density over its reference — but its phase inductance drops into the microhenry range, which produces large current ripple and harmonic distortion unless the drive switches fast or the phase is deliberately loaded with inductance. Low inductance is not low efficiency; it is a different drive problem. The same family of designs has demonstrated 96 percent efficiency at 19 N·m and 2,100 rpm with a wide-bandgap inverter.

Thermally, axial flux has an underrated advantage that rarely makes the brochure: the end windings are exposed on the disc faces, close to the housing, with short conduction paths. That is exactly the geometry that makes a joint survive continuous torque rather than a three-second peak — the distinction that separates a demo from a shift.

Cost four: the supply chain is still the binding constraint

There is no catalogue shelf of axial-flux joint motors in humanoid sizes. Small-diameter units require custom stator segmentation, custom concentrated-coil winding, multi-pole Halbach magnet rings with degree-level pole-arc control, and custom gap-parallelism gauges. Automotive-scale production has proven the manufacturing route is real, but automotive diameters, speeds and duty cycles are not knee-joint duty cycles.

And the installed base is the tell. As of October 2026, no major humanoid manufacturer has publicly confirmed axial-flux motors in a production joint. The evidence consists of supplier roadmaps, media reports about supply relationships, and a wave of Chinese joint-motor launches at trade fairs. Radial-flux frameless motors behind planetary or cycloidal stages remain the default for a boring reason: a century of manufacturing maturity, predictable second sources, and datasheets you can design against.

A decision table instead of a religion

None of this makes axial flux a bad choice. It makes it a specific one.

The pattern is simple enough to state as a rule: axial flux converts diameter into axial length, and charges the difference to inertia. Designs that need the length and can live with the inertia — a low-ratio knee, a compact hip, a legged robot with short legs — get real value. Designs that keep a high ratio and a big disc are paying twice for the same decision.

How Ti5ROBOT specifies a joint motor

We build integrated joint modules and humanoid actuators, and the motor topology is the last thing we choose, not the first. The sequence that works is: duty cycle, then continuous torque, then the axial budget, then the ratio, and only then the machine.

That ordering exists because the numbers interact. A motor chosen for peak torque at 33 mm will fail a continuous-torque test if its thermal path is an afterthought, and a joint chosen for a 100:1 ratio will feel dead in force control no matter how good the motor is. Our humanoid robot actuators and lightweight joint modules are specified this way, and the 30–40 frame family shows what the axial-budget exercise looks like at the small end, where every millimetre competes with the encoder and the harness.

Three questions separate a real axial-flux offer from a slide:

• At what thickness, mass and duty cycle was that torque measured? Peak torque at 20 °C for three seconds is a different product from 45 Nm continuous at 40 °C ambient.
• Which motor is the comparison baseline? "Two to four times" needs a named machine, the same cooling, and the same measurement point.
• How is the gap held? Bearing arrangement, housing stiffness in N/μm, gap tolerance in μm, and the measured axial runout under load.

If a supplier answers all three with numbers, the architecture is worth a prototype. If the answers are adjectives, it is worth a pass.

What to watch

Three developments would make the axial-flux arithmetic better without changing the physics. Automotive volume production spilling into robot-sized diameters would collapse the custom-tooling premium. Soft magnetic composites and PCB stators maturing would make the loss and drive problems routine rather than bespoke. Wide-bandgap drives becoming standard in joint controllers would neutralize the low-inductance penalty and let ironless designs compete on efficiency instead of just on thickness.

None of that changes the trade. Diameter buys axial length; inertia sends the bill. The joints that win will be the ones whose designers did the multiplication before the trade show, not after.

Source: IEEE/ASME Transactions on Mechatronics (2026), axial-flux motor–cycloid actuator optimization for wearable robotics; arXiv 2509.23561 (2025), high-torque-density PCB axial-flux motor for micro robots; Applied Sciences 15(6):3327 (2025), axial-flux PM motor design and verification; ICIRA 2023, biped actuator based on an axial-flux motor; Steinmetz core-loss formulation; industry disclosures at CIIF 2026 and humanoid actuator market surveys, October 2026.