Harmonic Reducer

What is a Harmonic Reducer

A harmonic reducer is a gear speed reducer that exploits the controlled elastic deformation of a thin metal ring to transmit motion, rather than the rigid meshing of conventional gear trains. Its formal name is strain wave gearing, and its best-known trade name is Harmonic Drive. The mechanism reduces input speed and multiplies torque between a fast input shaft, usually a servo motor, and a slow, high-torque output, usually a robot joint. What sets it apart from a planetary or worm reducer is that it reaches a single-stage ratio of roughly 30:1 to 320:1 with essentially no backlash, in a flat, lightweight package whose size and weight do not change with the ratio chosen.

The reducer consists of three core parts. The wave generator is an elliptical steel hub fitted with a thin, flexible ball bearing, and it is the input. The flexspline is a non-rigid, thin-walled steel cup with external teeth cut on its open rim, and it is typically the output. The circular spline is a rigid internal-gear ring with two more teeth than the flexspline, usually held fixed to the housing. When the wave generator turns inside the flexspline, it forces the flexspline into an elliptical shape so its teeth engage the circular spline at the two ends of the ellipse major axis, and disengage at the minor axis.

The history is well documented. American engineer C. Walton Musser invented strain wave gearing in 1957 and was granted US Patent 2,906,143 in 1959, and the technology was first licensed under the Harmonic Drive name in the United States and then in Japan. Through the 1970s and 1980s it became the enabling reducer for the first generation of precision industrial robots, because no other compact mechanism combined zero backlash with a 100:1 reduction. NASA adopted it for spacecraft mechanisms, and harmonic drive actuators served as wheel drives and antenna gimbals on the Mars Exploration Rovers Spirit and Opportunity, where mass, vacuum operation, and long maintenance-free life were decisive.

The application scale today is dominated by robotics. A six-axis industrial robot typically uses harmonic reducers in its three or four lighter wrist and forearm axes, where low inertia matters, and heavier RV cycloidal reducers in the base, shoulder, and elbow. Collaborative robots, which prize low weight and a hollow shaft for cable routing, use harmonic reducers in nearly every joint. Beyond robotics, harmonic reducers appear in semiconductor wafer-handling robots that must move without shedding particles, in CNC rotary tables and indexers, in medical imaging gantries, in optical and antenna positioners, and in aerospace flight-control and solar-array actuators.

Four engineering metrics decide whether a harmonic reducer is fit for a joint: the reduction ratio, the family of torque ratings, the positional accuracy and lost motion, and the rated service life under the actual duty cycle. These are not independent: a higher ratio raises output torque but lowers the allowable input torque, and running near peak torque or above the rated input speed shortens life sharply. The remainder of this guide decodes each of these so that a procurement engineer can read a component-set datasheet without translation.

Structure and Operating Principle

The strain wave principle is the single most important concept to grasp before reading any spec sheet, because every headline figure, zero backlash, high ratio, high stiffness, follows directly from it. The reducer transmits torque through the controlled bending of the flexspline, not through the conventional rolling contact of rigid gear flanks. The elliptical wave generator deflects the flexspline so that its external teeth fully mesh with the internal teeth of the circular spline at the two ends of the major axis, are fully disengaged at the minor axis, and are partially engaged in between.

Because the flexspline has two fewer teeth than the circular spline, the geometry produces the reduction. For every full revolution of the wave generator, the zone of tooth engagement sweeps once around the ring, and the flexspline is forced to lag the circular spline by exactly the two-tooth difference. If the circular spline is held fixed, the flexspline therefore rotates slowly in the direction opposite to the wave generator, advancing by two teeth per input turn. A flexspline with 200 teeth meshing a 202-tooth circular spline advances 2 teeth out of 200 per input revolution, giving a 100:1 reduction.

This mechanism is the source of three signature properties. First, zero backlash: at all times the teeth are preloaded into engagement at both ends of the major axis, so there is no clearance to traverse when the load reverses. Second, high torque density: because up to 30 percent of the total tooth count is in mesh simultaneously, far more than the one or two teeth that carry load in a spur or planetary mesh, load is shared across many teeth, raising both torque capacity and torsional stiffness. Third, smooth, averaged motion: with so many teeth engaged, individual tooth errors are averaged out, giving high positional repeatability.

The price of bending a steel ring twice per input revolution is fatigue and heat. The flexspline experiences a fully reversed bending stress cycle on every wave-generator turn, so its fatigue life, expressed as an L10 rated life in hours, is a published, load-and-speed-dependent number rather than infinite. The repeated flexing also dissipates energy, which is why harmonic reducers run warmer than planetary gears and why efficiency and input speed are coupled. The standard structural variants and their trade-offs are summarized below.

One practical consequence of the principle deserves emphasis for selection. Because the ratio is fixed entirely by the flexspline and circular spline tooth counts, every ratio in a given size shares the same envelope, mounting, and bearing. Changing a robot joint from a 50:1 to a 160:1 reducer does not change the mechanical interface, only the input speed and output torque available. This is why component-set catalogs list one frame size with a column of selectable ratios, a structure that does not exist for multi-stage planetary trains.

Construction Types: Cup, Hat, Pancake

Although the strain wave principle is identical across the catalog, harmonic reducers are sold in three flexspline geometries, plus a distinction between bare component sets and complete housed units. Choosing the wrong geometry is a common and expensive selection error, because the cable-routing through-bore, axial length, and torque capacity differ substantially while the part numbers look similar. The table below compares the three mainstream geometries on the metrics that drive a mechanical layout.

Cup-type reducers, such as the CSF and the higher-torque CSG series, use a deep cup-shaped flexspline closed at one end and connected to the output at the cup bottom. This is the original geometry and remains the highest-efficiency, highest-torque-density option, which is why it is the default for servo gearheads and for compact robot wrists where no cable needs to pass through the center. The closed cup end, however, blocks the axis, so any wiring must route around the outside of the reducer.

Hat-type reducers, also called silk-hat from the shape of the flexspline, fold the flexspline rim outward into a brim that doubles as a clean mounting flange. The key benefit is a large hollow through-bore down the rotational axis, through which electrical cables, pneumatic lines, and even laser or coolant supply can be routed without a slip ring or cable carrier. This is decisive for collaborative robots and for any joint where internal cable routing improves reliability and appearance. The SHF and high-torque SHG series cover this geometry, at a small efficiency penalty versus an equivalent cup unit.

Pancake-type reducers, such as the CPU and FB families, omit the cup wall entirely and use a flat ring flexspline running between a fixed circular spline and a rotating dynamic spline of equal tooth count. The result is the shortest possible axial length and the largest central hole for its diameter, ideal where axial space is tightly constrained or a very large bore is needed. The trade-off is reduced torque capacity and lower efficiency than a cup unit of the same diameter, so pancake types are chosen for packaging reasons, not for maximum performance.

Cutting across all three geometries is the distinction between a component set and a housed unit. A component set supplies only the three gear elements, leaving the machine designer to provide bearings, housing, and the input-output interface, which minimizes mass and lets the reducer be integrated tightly into a custom joint, the norm in robot arms. A housed unit, sometimes called a gear unit or an actuator, adds a precision cross-roller output bearing, a sealed housing, and a defined input coupling, so it bolts directly to a motor and a load. Housed units cost more and weigh more but remove the design and assembly burden, which suits machine builders and lower-volume applications.

Ratio, Tooth Math, and Standards

The reduction ratio is the first number a buyer fixes, and it follows directly from the tooth math of Chapter 2. With the wave generator as input and the circular spline held fixed, the reduction ratio equals negative Nf divided by the quantity Nc minus Nf, where Nf is the flexspline tooth count and Nc is the circular spline tooth count. Because the standard tooth difference is two, the magnitude simplifies to Nf divided by 2, and the negative sign means the output rotates opposite the input. A 100:1 unit therefore has a 200-tooth flexspline and a 202-tooth circular spline.

This single-stage ratio range spans roughly 30:1 to 320:1, with the commercially common values being 50, 80, 100, 120, and 160, the set listed for the Harmonic Drive CSF and CSG cup-type series and mirrored by domestic series such as Leaderdrive LHD. Where ratios beyond 320:1 are needed, two harmonic stages or a harmonic stage following a planetary stage are used, but a single set covers most robot joints. The table below shows how the same frame size yields different ratios purely through tooth count, alongside the published repeated-peak torque envelope of the CSG cup-type series, which spans roughly 23 to 3,419 Nm across its ten sizes.

On the standards side, harmonic reducers in the Chinese market are governed by GB/T 14118, the national standard titled Harmonic Drive Reducers, which defines terminology, type designations, performance grades, and acceptance requirements. Its companion document GB/T 14623 specifies the test methods, including the endurance test in which the reducer runs continuous forward and reverse cycles with acceleration and deceleration at maximum speed under overload, and the noise checks used as a preliminary quality screen. Procurement specifications for Chinese-built robots routinely cite these two standards together.

Manufacturers also publish their figures against general gear-metrology conventions, so transmission accuracy, lost motion, and torsional stiffness are stated in the consistent units of arc-seconds, arc-minutes, and Nm per arc-minute. When comparing a domestic series to an imported one, the engineer should confirm that both quote the same defining condition, for example lost motion measured at plus-or-minus 4 percent of rated torque, because a favourable-looking number measured under a gentler condition is not comparable. The next chapter decodes each of these spec-sheet fields.

Key Specification Parameters

A harmonic reducer datasheet lists many fields, but selection turns on a compact set: the four torque ratings, ratcheting torque, transmission accuracy, lost motion, torsional stiffness, allowable input speed, rated life, and efficiency. Each is decoded below, with the typical magnitudes engineers should expect from a quality component set.

The four torque ratings are the heart of sizing, and confusing them is the most common error. Rated torque is the continuous torque the reducer can carry indefinitely at rated input speed; it sets the basis for the L10 life. Repeated peak torque is the higher torque permitted during acceleration and deceleration phases of a normal cycle. Momentary (maximum) peak torque is the still higher torque allowed only for emergency stops and rare overloads, a few times the rated value. Average torque is the duty-cycle-weighted figure that, with average speed, actually determines fatigue life. The four are listed separately precisely because they cannot be collapsed into one number.

Ratcheting torque is a failure-limit unique to strain wave gears: at a sufficiently high static overload, the flexspline teeth disengage and jump, or ratchet, over the circular spline teeth, permanently damaging the gear and losing the homed position. It is typically well above momentary peak torque, but it must bound the worst-case collision or jam load in a robot joint. Transmission accuracy is the deviation between actual and theoretical output angle over one output revolution; quality harmonic gears achieve on the order of tens of arc-seconds, and selected high-precision units reach within 10 arc-seconds, a figure no comparably compact reducer matches.

Lost motion is the small angular play measured at the output when a low reverse torque, conventionally plus-or-minus 4 percent of rated torque, is applied with the input locked. Harmonic Drive specifies lost motion under 1 arc-minute, maintained for the life of the gear, which is what the marketing term zero backlash means in practice. Torsional stiffness describes how much the output deflects under load and is published as a three-region curve, often summarized by the K3 stiffness value in Nm per arc-minute; it matters for servo loop tuning and for settling time, not just for static accuracy.

Allowable input speed distinguishes a continuous limit from a momentary limit and is size-dependent, for example near 5,600 rpm continuous and 7,500 rpm maximum for a mid-size unit, lower for larger frames. Because the flexspline flexes twice per input turn, heat and efficiency loss rise steeply above roughly 3,500 rpm, so the average input speed over the cycle, not the peak, governs thermal rating. Rated life is the L10 fatigue life in hours, commonly 7,000 to 10,000 hours, computed from average torque, average input speed, and a duty factor. Efficiency is typically about 70 to 90 percent, higher at higher load and lower ratio, with a representative figure above 80 percent for a 100:1 unit at rated torque and speed; it falls at light load, low temperature, and very high input speed.

Selection Decision Factors

To convert the preceding five chapters into a chosen part number, follow the sequence below. Most selection failures come not from one wrong field but from deciding torque before the duty cycle is known, or from ignoring the thermal and life consequences of the chosen ratio. These eight steps form a fixed RFQ template for a harmonic reducer.

1. Define the duty cycle first: map the motion profile into continuous load torque, acceleration and deceleration torque, emergency-stop torque, average torque, and average input speed. Every subsequent torque check refers back to these numbers, so guessing them invalidates the selection.
2. Choose ratio and frame size together: the ratio sets output torque and the reflected inertia the motor sees, while the frame size sets the torque envelope. Pick the smallest frame whose rated torque exceeds your continuous load torque and whose repeated peak exceeds your acceleration torque.
3. Verify all four torque limits: continuous load below rated torque, acceleration below repeated peak, e-stop below momentary peak, and the worst-case collision below the ratcheting torque. A single violated limit disqualifies the part.
4. Confirm rated life: compute the L10 life from average torque and average input speed and compare it to the required service hours with margin. A unit that passes every torque check can still fail if the average load shortens life below the maintenance interval.
5. Select the construction type: cup for maximum efficiency and compactness, hat for a hollow cable bore, pancake for minimum axial length. Decide component set versus housed unit based on whether you will provide your own output bearing and housing.
6. Check accuracy and stiffness against the application: match transmission accuracy and lost motion to the positioning tolerance, and confirm the torsional stiffness K3 supports the required servo bandwidth and settling time. Over-specifying accuracy raises cost without benefit.
7. Respect the speed and thermal envelope: keep average input speed within the continuous limit and below the ~3,500 rpm efficiency knee where possible, and confirm the ambient temperature and duty factor used in the life calculation match reality. Add cooling or derate for continuous high-speed duty.
8. Confirm interface and supply: output bolt pattern, integrated cross-roller bearing moment and axial-radial load ratings, input coupling, lubrication (grease versus oil) and relubrication interval, plus the maker's lead time, spare availability, and field support.

One dimension routinely underweighted at the quotation stage is serviceability and supply security. A harmonic reducer is a wear part with a finite L10 life, so spare-part lead time and the availability of compatible cross-brand replacements determine line uptime years after purchase. Harmonic Drive SE and Harmonic Drive Systems hold the original patents and supply premium robotics and aerospace with the CSF, CSG, SHF, SHG, CSD, and CPU series; Sumitomo offers strain wave gearheads. Among domestic suppliers, Leaderdrive (LHD, LHS, LCS, LHSG series) and Laifual have reached backlash and accuracy parity for standard duty cycles at roughly 70 to 85 percent of the Japanese benchmark on torque density and life, while pricing at 40 to 60 percent, which suits cost-driven and high-volume programs. Finally, do not confuse suppliers: Nabtesco is the dominant maker of RV cycloidal reducers, not harmonic drives.