Radial flux: what is inside almost every joint today

In a radial-flux motor the rotor and stator are concentric cylinders, and the magnetic field crosses the thin cylindrical gap between them pointing outwards from the shaft. The frameless torque motors in many humanoid arms and the large outrunner motors in quasi-direct-drive legs are both of this type; the MIT Mini Cheetah actuator, the template for many low-ratio legged joints, wraps a single 6:1 planetary stage inside a radial outrunner. The layout is mature, cheap to wind in volume and supported by a deep catalogue of off-the-shelf stators and rotors.

Its weakness for a robot joint is geometry. At fixed electrical and magnetic loading, radial-flux torque scales with the rotor diameter squared times the stack length. A designer who wants more torque either makes the motor longer, which adds width to a hip or knee, or makes it wider and accepts that the useful active area only grows with the square of the diameter.

Axial flux: turn the gap sideways

An axial-flux motor puts the air gap in a plane perpendicular to the shaft. The rotor is a steel disc carrying a ring of magnets; the stator is a disc of coils facing it. Flux runs parallel to the shaft, which is why these machines are also called axial-gap or pancake motors. A common robot layout puts one stator between two rotors, so that the magnetic pull on either side of the stator largely cancels and the stator needs no back iron of its own.

The scaling changes too. The active area is an annulus whose area grows with the diameter squared, and the force on it acts at a radius that grows with the diameter, so at fixed thickness torque rises roughly with the cube of the diameter. The authors of a 2025 PCB axial-flux motor paper cite exactly this scaling as the reason the topology suits low-speed, high-torque robot joints, alongside the flat shape that matches a disc-shaped joint envelope and gives coils a short path to shed heat.

Why the shape matters in a humanoid

Humanoid joints are short and wide. A hip or knee actuator stacks a motor, a reducer, an output bearing, one or two encoders, a brake on some joints and a servo drive into a package that has to fit across the width of a pelvis or a thigh. Every millimetre of motor length is a millimetre added to that stack. A flat motor that delivers its torque from diameter rather than length frees axial space for the reducer and electronics, or makes the leg narrower.

Higher motor torque can also mean a lower gear ratio. Hong Kong University of Science and Technology (Guangzhou)’s AMDT Lab describes its back-to-back axial-flux robot motor as a way to raise torque density so that the reduction ratio of the precision reducer can come down. A lower ratio means less reflected inertia, a joint that is easier to back-drive and a reducer that is cheaper to make.

What research prototypes show

No major humanoid maker publishes axial-flux joint data, so the evidence comes from research actuators. The most humanoid-relevant is a Zhejiang Lab actuator presented at ICIRA 2023. It pairs an axial-flux motor with 12.5 N·m peak torque with a two-stage planetary reducer of about 16:1, reaching 206 N·m at the output with a maximum speed of 120 rpm in a 2 kg package, cooled by a small fan. The team reports that it drives a biped that walks at 6 km/h.

At the other end of the scale, a 2025 paper describes a 19 mm, 25 g PCB-stator axial-flux motor for micro-robots, with 158 mN·m stall torque: 25% more than a commercial reference motor in about 25% less volume, a 62% gain in stall torque density. The same prototype peaks at 60% efficiency against the reference motor’s 83%, which is a reminder that density figures alone do not tell the whole story.

Thin is not the same as dense

A 2026 Harvard study of wearable ankle actuators gives one of the few published side-by-side comparisons. Its dual-rotor PCB axial-flux motor with a 26:1 cycloid reducer delivers 45 N·m peak from about 600 g in a 33 mm thick package, or 75 N·m per kg. Two radial-flux designs in the same table beat it on torque density, one at 120 N·m per kg. The axial design’s advantage is thickness: 33 mm against about 75 mm and 63 mm for two of the radial designs. Only one radial design, at 24 mm, is thinner, and it uses a 200–425:1 reduction to get there.

That is the realistic case for axial flux in humanoid joints: not a universal jump in N·m per kg, but more torque per millimetre of joint width, which designers can spend on a narrower leg, a lower gear ratio or a bigger bearing.

Why most joints still use radial flux

• Magnetic pull across a tiny gap. Rotor magnets pull hard on the stator iron across a gap of well under a millimetre (0.25 mm in the PCB motor above). Bearings and housings have to hold that gap flat under load and impact, which is one reason dual-rotor layouts are common.
• Rotor inertia. A wide rotor has more inertia than a slim one, and in a geared joint that inertia is multiplied by the square of the ratio at the output. Axial flux helps most when it lets the ratio come down.
• Cramped inner radius. Coil segments narrow towards the centre, leaving less room for copper and connections where the hollow shaft, bearings and cables also want space.
• Supply chain. Volume production of axial-flux motors is real in cars, with Mercedes-Benz producing YASA-designed motors in Berlin since 2026, but standard small robot sizes are scarce. China’s Panhu Power says its 300,000-unit-a-year line will serve humanoids, without naming a robot customer.

What to ask an axial-flux joint supplier

Ask for continuous and peak torque at the actuator output, the motor-only torque constant, total actuator mass and thickness, rotor inertia, and the reference motor behind any percentage gain. Ask how the air gap is held under a fall load, and whether the published figures come from test or simulation: of the five designs in Table 1, the robot-dog actuator’s figures come from finite-element analysis.

FAQ

Sources