One lobe fewer: how the drive works

Three parts do the work. A ring of hardened pins sits fixed in the housing. Inside it runs a disc whose rim is a smooth wave of rounded lobes, mounted on a bearing over an eccentric on the motor shaft. Through holes in the disc pass output pins, bolted to the output flange.

Turn the motor and the eccentric swings the disc round in a small circle, pressing its lobes against the pins. Because the disc has one lobe fewer than the ring has pins, it can’t simply orbit: each time the eccentric goes round, the disc rolls one lobe further along the ring and ends up turned backwards by one lobe pitch. With 11 pins and 10 lobes, ten input turns give one output turn, in the opposite direction. The general rule is the lobe count divided by the difference between pins and lobes, which with the usual one-lobe gap is just the lobe count.

The output pins take the slow rotation and ignore the wobble. Each hole in the disc is larger than its pin by exactly twice the eccentricity in diameter, so the pin rolls round the inside of the hole as the disc orbits and only the turning motion reaches the flange. The idea is old: the German engineer Lorenz Braren patented the cycloid gearbox in the 1920s and founded Cyclo GmbH in Munich in 1931; licensed production in Japan, where Sumitomo still sells the Cyclo line, started in 1939.

Rolling contact, and why it takes a beating

What sets the cycloid apart is how load is shared. In a planetary gear, torque passes through a few involute teeth at a time. In a cycloidal drive the lobes roll on pins, usually with roller sleeves, and many of them carry load at once. Nabtesco says nearly half the cams on each disc are in contact with the rollers at any moment, and with two discs phased 180° apart the contact covers almost the full circle.

That load sharing shows up in the catalogues as shock capacity. Nabtesco rates its RV gears to absorb up to 500% of rated torque in an emergency stop; the RV-80N is rated at 784 N·m with a momentary maximum of 3,920 N·m. Spinea lists the same 5× ratio for its small TwinSpin units, and Sumitomo’s smallest Fine Cyclo DA pairs 257 N·m nominal with 1,225 N·m momentary peak. Lost motion in all three ranges is under one arc minute, and Nabtesco quotes pure backlash of 0.1 to 0.2 arc minutes.

For a humanoid that matters most in the legs. A fall or a hard landing pushes torque back into the joint, multiplied by the reflected inertia of the motor. A strain-wave gear relies on a thin flexspline to survive that; a cycloid relies on rigid steel lobes in compression.

The price: imbalance, precision and weight

Nothing comes free. The disc runs off-centre, so one disc on its own shakes the whole drive. Two discs phased 180° apart cancel the static imbalance but leave a small dynamic one, which is why high-speed designs use three or more. The research actuator described below uses two discs for its even 10:1 ratio and would have needed three for an odd one.

Then there’s machining. The clearance between lobes and pins is tiny, and any error in the profile turns straight into backlash or torque ripple. The 10:1 research actuator had its discs and rings wire-EDM cut in 4140 steel to ±2 µm. Shrinking the design to a few tens of millimetres, as Chinese suppliers such as Siggear now advertise, makes every micron count more.

Efficiency is decent rather than outstanding. Single-stage cycloids can approach 93% and two-stage units about 86%; Nabtesco states up to 85% for the RV-N. The research actuator measured 82.3% peak efficiency from electrical input to mechanical output, motor losses included.

Weight is the real catch for humanoids. Industrial RV gears were built for the base axes of welding and handling robots, and it shows: the smallest RV-N, the RV-25N, is 133 mm across and weighs 3.8 kg for 245 N·m rated. That is a strong torque-per-kilogram figure, but the absolute size is closer to a car-factory robot’s wrist than to a humanoid knee.

RV: the two-stage industrial version

Nabtesco’s RV design puts a spur-gear stage in front of the cycloid. The input pinion drives three spur gears, each on an eccentric shaft with two eccentrics, and those shafts move two cam discs inside a ring of rollers. The first stage takes some of the reduction and the three crankshafts share the load, so the combined ratio runs from about 30:1 to over 300:1 with very high tilting stiffness. Nabtesco says bending moment capacity is about 250% higher than conventional planetary gearboxes of comparable size.

Nabtesco is the incumbent, and market researchers now track a small “miniature RV for humanoids” segment. Intel Market Research puts it at USD 34.4 million in 2024 and credits Nabtesco with about 28% of it, figures we haven’t been able to verify. Slovakia’s Spinea, maker of the TwinSpin, is now part of Timken.

Cycloids in legged-robot joints

The more interesting work for humanoids uses the cycloid at low ratios, as a tougher alternative to the planetary gears in quasi-direct-drive actuators. Rafeedi and co-authors built a 10:1 cycloidal QDD actuator weighing 1.40 kg, with 37.5 N·m continuous and 89.9 N·m measured peak torque, the peak capped by the motor driver’s current limit rather than the gear. They estimate a matching compound planetary would need six planets to carry 120 N·m. It back-drives with 1.99 N·m static and 1.36 N·m dynamic torque, has 7.0 arc minutes of backlash and reaches 34.3 Hz torque-control bandwidth.

On the commercial side, PHYBOT (listed by 36Kr as Dongyi Technology) showed its PhyArc family of integrated cycloidal joint modules at WAIC 2026 alongside its M1 humanoid. The PhyArc 78 combines a 25:1 cycloid with a torque motor, drive and dual encoders in a 78 by 79 mm package: 40 N·m rated, 136 N·m peak, 0.98 kg, on a 24 to 72 V supply, listed at USD 2,050. Korea’s Bonsystems pitches a pinless cycloidal reducer for humanoid joints, replacing the pin ring with a machined profile to cut part count. None of these has yet displaced the strain-wave gear in a volume humanoid.

What to ask a cycloidal supplier

• Which torque is which? Rated, acceleration, momentary and E-stop figures differ by up to 5×, and actuator “peak” is often a motor-current limit.
• How many discs, and how is the remaining imbalance handled at the input speeds your joint will see?
• Lost motion and backlash measured at what preload, and how they change after a life test.
• Profile tolerance and how it’s made – ground, wire-EDM or sintered – since that sets ripple, noise and price.
• Back-drive torque, if the joint has to yield to contact or be moved by hand.

FAQ

Sources