Three ways to turn 3,000 rpm into a knee

A frameless motor in a humanoid joint wants to spin fast and produce little torque. The limb wants the opposite. The reducer sits between them and trades speed for torque, and the way it does that trade decides almost everything else about the joint: how stiff it is, how much it heats up, whether the controller can feel a collision through motor current, and whether a fall on concrete cracks it.

A strain-wave drive, sold by Harmonic Drive Systems under the Harmonic Drive name and by dozens of Chinese makers as a harmonic reducer, uses an elliptical wave generator that flexes a thin steel cup, the flexspline, against a rigid circular spline with two more teeth. Each turn of the ellipse walks the cup back by two teeth. That gives ratios of 50:1 to 160:1 in a single stage, essentially zero backlash, and a package thin enough to disappear inside a wrist. A planetary reducer is the ordinary gear train: a sun gear drives several planets inside a ring, and each stage manages 3:1 to 10:1, so you stack stages for more. A cycloidal reducer uses an eccentric cam to roll a lobed disc around a ring of pins; many lobes carry the load at once, which is why it shrugs off impacts that would wreck a flexspline.

The numbers engineers actually compare

Joint-module maker HONPINE publishes a side-by-side that matches what most datasheets show. Harmonic drives hold backlash to a single arc-minute or less and manage 80 to 90 percent efficiency, but the flexspline limits them to two to three times rated torque in a shock and to roughly 7,000 to 10,000 hours of life. Planetary stages run at 90 to 97 percent efficiency and last 20,000 to 30,000 hours, at the price of 3 to 10 arc-minutes of play. Cycloidal units sit in between on backlash and efficiency yet lead on toughness: five to ten times rated torque in overload and more than 50,000 hours of service. Nabtesco, which builds the RV cycloidal reducers in most large industrial robots, puts pure backlash of its cycloidal units at 0.1 to 0.2 arc-minutes and notes that strain-wave gears top out at 382 N·m, a ceiling set by the elastic flexspline.

Why humanoids split into two camps

Industrial arms settled this question decades ago: harmonic in the wrist, cycloidal RV in the base, and HONPINE reckons more than 90 percent of robot joint modules today still use either a harmonic or a planetary reducer. Humanoids reopened it, because a walking robot is a different machine. It falls over. It slams its heel into the floor a few thousand times an hour. And its controller needs to feel the world, not just position a tool.

That last requirement created the quasi-direct-drive camp. A QDD joint pairs a large-diameter, high-torque motor with a planetary stage of only 5:1 to 20:1. The low ratio keeps reflected inertia small and efficiency high, so the joint can be pushed from the outside and the motor current tells the controller how hard. Figure 02, the electric Atlas and Agibot’s Q1 are all described as QDD designs, and Unitree builds its own low-ratio planetary actuators for the G1 at a reported average cost of around $300 per unit. The cost of that compliance is torque density: with little gear reduction, the motor has to be big and heavy, and it heats up when the joint holds a static load.

The harmonic camp takes the opposite bet. A 100:1 strain-wave joint gets far more torque from a small motor, holds position without current, and packs into a wrist or forearm where there is no room for a large rotor. The trade is that you cannot feel through it: high ratio plus the flexspline’s own friction makes the joint nearly impossible to back-drive, so builders add a joint torque sensor, which is roughly 30 percent of a rotary actuator’s cost in HONPINE’s own breakdown. Tesla’s Optimus is the visible example, with something on the order of 14 strain-wave reducers in its rotary joints alongside the 14 roller-screw linear actuators covered in our linear vs rotary explainer. Chinese reporting on the Optimus supply chain also places planetary roller screws in the heavy lower-limb joints, so even one robot mixes rotary and linear actuation.

• Pick harmonic when the joint must be compact and precise, holds loads for long periods, and can carry a torque sensor: wrists, shoulders, waists, most arm joints on manipulation-first robots.
• Pick a low-ratio planetary when the controller relies on proprioceptive force sensing and the robot must survive contact, falls and fast running: hips, knees and ankles on locomotion-first and low-cost robots.
• Pick cycloidal when the joint sees repeated shock, needs stiffness against tilting loads and must run for tens of thousands of hours: heavy industrial humanoids and load-bearing hip and ankle joints, where Bonsystems, Nabtesco and Nidec are now pitching compact pin-less designs.

Who makes the reducer

Strain-wave gearing was a Japanese near-monopoly for forty years. Harmonic Drive Systems is cited as holding as much as 85 percent of the strain-wave segment in 2023, and Nabtesco around 60 percent of RV cycloidal reducers for medium and large industrial robot joints. Humanoids are eroding the first number fast. Chinese producers went from about 15 percent of the strain-wave market in 2022 to roughly 38 percent by early 2024, led by Leaderdrive, which grew revenue 47 percent to 570.7 million yuan in 2025, counts Agibot and UBTech as customers, and has been reported as a validated reducer supplier to Tesla. Green Harmonic claims more than 60 percent of the domestic Chinese harmonic market at 40 to 60 percent of Harmonic Drive’s prices. Shuanghuan Driveline is adding a Suzhou plant with 500,000 units of annual reducer capacity in 2026.

The planetary side looks different because there is less to protect. Gears are a commodity, so QDD builders such as Unitree and Figure design and often machine their own stages, and Tier-1 automotive suppliers see an opening. Schaeffler’s production-ready humanoid actuator, shown at CES 2026, is a two-stage planetary unit rated 60 to 250 N·m with motor, encoder and controller in one housing. Nidec, meanwhile, is the one supplier hedging all three: its FLEXWAVE strain-wave, KINEX cycloidal and planetary lines were pitched together at Automate 2026 as options for the same humanoid joint.

What it does to the bill of materials

Reducers are not the single biggest line on a humanoid BOM, but they are the one that decides who gets paid. 36Kr’s breakdown of a mass-production Optimus targeting under $20,000 puts reducers at around 15 percent, or roughly $3,000 per robot, behind actuator assembly at 35 percent and lead screws at 18 percent. Next Financial’s estimate is that a humanoid carries somewhere between 15 and 40 precision reducers against a handful in a conventional arm, which is the arithmetic behind every reducer stock rally in Shanghai this year. Global humanoid shipments rose nearly 480 percent to 13,318 units in 2025; at even 15 reducers each, that is a market measured in the hundreds of thousands of units, on its way to Omdia’s projection of 2.6 million robots a year by 2035.

Price pressure runs in one direction. Chinese harmonic units already sell at half the Japanese price, QDD modules are quoted at $800 to $3,000 depending on torque rating, and Unitree has shown what happens when a builder machines its own planetary stages. The likely outcome is not one winner but a hardening of the split described above: cheap, in-house planetary joints in the legs of volume robots, imported or licensed strain-wave units in the arms of precision robots, and cycloidal creeping into heavy-duty industrial humanoids from the bottom up.

FAQ

Sources