Mechanical

Harmonic Drive (Strain Wave Reducer)

An elliptical wave flexes a cup of teeth into a ring of teeth — once per input turn, ratio N/2

A harmonic drive uses an elliptical wave generator to flex a thin-walled flexspline into a rigid circular spline with two more teeth, producing huge ratios in a flat package with near-zero backlash. The reduction ratio equals N/2, where N is the flexspline tooth count. Standard in industrial robot wrists, satellite drives, and precision indexers.

  • Single-stage ratio30:1 to 320:1
  • Backlash< 30 arc-seconds (≈ 0.008°)
  • Tooth-count difference2 (circ_teeth − flex_teeth)
  • Three componentsWave generator, flexspline, circular spline
  • Efficiency60 to 85% (ratio-dependent)
  • InventorC. Walton Musser, 1957 (US Pat 2,906,143)

Interactive visualization

Press play, or step through manually. The visualization is yours to drive — try it before reading on.

Open visualization fullscreen ↗

Watch the 60-second explainer

A condensed visual walkthrough — narrated, captioned, under a minute.

How a harmonic drive works

A harmonic drive has only three working parts:

  • Wave generator — an elliptical steel cam mounted to the input shaft, wrapped in a thin ball bearing. Spinning the cam makes the bearing race trace an ellipse that orbits with the input.
  • Flexspline — a thin-walled steel cup with external gear teeth cut into its open end. The cup flexes elastically into the ellipse imposed by the wave generator.
  • Circular spline — a rigid internal-toothed ring with exactly two more teeth than the flexspline.

The geometric trick: at the long-axis ends of the ellipse, the flexspline teeth fully engage the circular-spline teeth; at the short-axis ends they fully disengage. As the wave generator rotates, the engagement zone walks around at input speed. But because the flexspline has two fewer teeth than the circular spline, after one full rotation of the wave generator the flexspline has been forced to step backwards by exactly two teeth relative to the circular spline. That two-tooth-per-input-revolution lag is the entire reduction.

Reduction ratio (with circular spline fixed, flexspline output):

  i = N_flex / (N_circ − N_flex)

For the standard 2-tooth difference (N_circ = N_flex + 2):

  i = N_flex / 2

A 100-tooth flexspline against a 102-tooth circular spline gives 50:1.
A 200-tooth pair gives 100:1.   A 320-tooth pair gives 160:1.

Two consequences fall out of this: (1) reduction ratio scales with tooth count, so harmonic drives are not "small" gear-tooth devices — a 100:1 unit literally has 200 teeth on its flexspline, and 50+ of those teeth share load at any moment. (2) The output rotates opposite the input when the circular spline is fixed; flipping which member is fixed changes the sign and slightly the magnitude of the ratio.

Worked example: a robot-wrist reducer

Take a typical industrial 6-axis robot's wrist joint, where harmonic drives are universal:

Flexspline tooth count:       N_flex = 200
Circular spline tooth count:  N_circ = 202   (difference = 2)
Reduction ratio:              i = 200 / 2 = 100:1
Servo motor:                  3000 RPM rated, ~3 N·m rated torque
Output speed (max):           3000 / 100 = 30 RPM = 180°/s
Output torque (rated, η=80%): 3 × 100 × 0.80 = 240 N·m

Backlash spec:    ≤ 30 arc-seconds (0.0083°)
At 0.5 m link length, that's ≤ 0.072 mm of tip drift per direction-reversal.
Repeatability of a typical 6-axis arm:    ±0.05 mm at end effector.

The numbers explain why harmonic drives are the default for the smaller wrist joints (J4, J5, J6) of every major industrial-robot brand: 240 N·m of stiff, backlash-free output torque from a 3 N·m motor, in a pancake unit ~80 mm in diameter and ~30 mm thick. Try to do that with a planetary gearbox and you're looking at three stages of reduction, more axial length, and ten times the backlash.

Real-world specs and applications

ApplicationTypical ratioBacklashNotes
Industrial robot wrist (ABB IRB, KUKA KR, Fanuc M-series)80:1 to 160:1≤ 30 arc-secStandard component; OEM is usually Harmonic Drive Systems Inc. (HDSI) or Nidec
Collaborative robot joints (UR, Franka, Doosan)50:1 to 100:1≤ 30 arc-secHollow-shaft variant carries cabling through the joint center
Satellite solar-array drive200:1 to 320:1≤ 1 arc-minVacuum-rated, dry-lubricant flexspline (MoS₂ coated)
Semiconductor wafer-handling stage30:1 to 80:1≤ 1 arc-minCleanroom-rated, sealed grease, sub-micron repeatability at chuck
Telescope tracking drive (RA / Dec)160:1+≤ 5 arc-sec (selected unit)Sidereal tracking demands minimal cogging
Antenna pointing (military, broadcast)100:1 to 200:1≤ 30 arc-secOften back-to-back duals for stiffness
Camera gimbal (cine, surveillance)50:1 to 100:11 to 3 arc-minCost-sensitive, smaller cup diameter, lighter alloy

Harmonic vs other precision reducers

Harmonic (strain wave)PlanetaryCycloidal (RV-style)WormSpur/helical reductionDirect-drive (no reducer)
Single-stage ratio range30:1 to 320:13:1 to 10:1 (per stage)30:1 to 200:15:1 to 100:13:1 to 5:1 per stage1:1
Backlash (typical)≤ 30 arc-sec5 to 15 arc-min≤ 1 arc-min5 arc-min to 1°10 to 30 arc-min0 (none)
Torsional stiffnessModerateHighHighestModerateModerateDetermined by motor
Shock load toleranceModerate (ratchets out)HighHighestModerateHighExcellent
Efficiency (typical)60 to 85%92 to 97%85 to 92%50 to 90%95 to 99%~100% (motor only)
Axial length (relative)Pancake (very thin)LongModerateModerateLongWhatever the motor is
Cost (relative)3 to 5×4 to 8×0.5 to 1×1 to 2×3 to 10× (large motor)
Typical homeRobot wrists, satellites, indexersServo speed reducers, carsRobot main joints, pressesConveyors, jacks, liftsGearboxes, vehiclesTurntables, frameless servos

Variants: cup, hat, pancake, dual-stage

  • Cup-type (the original Musser geometry). The flexspline is a closed-bottom cup with the input shaft passing through the open end into a small wave-generator hub. This is the standard form most robot wrists use.
  • Hat-type (top-hat). A reversed-flange variant where the flange is on the open end of the cup. The flange is the output mounting surface, putting the output coplanar with the housing face. Useful when axial length is critical.
  • Pancake (ring-type). Removes the cup bottom entirely — the flexspline becomes a thin ring. Loses some torsional stiffness but creates a hollow-shaft pass-through for cables, hoses, or laser beams. Standard on collaborative robots where signal and power cabling needs to route through the joint.
  • Dual-stage / compound. Two flexspline-circular-spline pairs share one wave generator; one circular spline is fixed, the other rotates. Lets the ratio reach the thousands (e.g. 1000:1 in a single compact unit) at the cost of efficiency.
  • Component sets vs sealed units. Component sets (just wave generator + flexspline + circular spline) are integrated into a customer's own housing — common in robotic and aerospace builds. Sealed units (with their own crossed-roller output bearing and lubricant fill) drop straight into a machine. The sealed-unit market is dominated by Harmonic Drive LLC and a handful of Asian competitors (Nidec, Leaderdrive, Laifual).

When to use a harmonic drive

  • You need 50:1 or more reduction in a single stage — multi-stage planetary gets bulky fast at these ratios.
  • Backlash must be effectively zero — robot end-effector positioning, telescope tracking, semiconductor stages.
  • Axial length is at a premium — pancake form factor outperforms any other reducer on this metric.
  • Cabling or laser pass-through is required — hollow-shaft variants accept up to 70% of the unit OD as a clear bore.
  • The duty cycle is moderate, not 24/7 extreme overload — flexspline fatigue life caps the application.

Pick a different reducer when shock loading is severe (cycloidal RV-style is the robot-base alternative), when efficiency is paramount (planetary stages run 92-97%), when cost is the limiter (a worm or planetary is cheaper), or when extreme stiffness matters more than zero backlash.

Common failure modes and pitfalls

  • Flexspline fatigue cracking. The cup wall flexes elastically through tens of millions of cycles. Cracks initiate at the open-end rim where bending stress and tooth-root stress superpose. Surface defects (machining marks, corrosion pits) accelerate initiation. Once a crack starts it propagates back toward the closed end and the unit ratchets, then loses position. Inspection method: borescope the flexspline ID looking for surface cracks; replace at the first hint.
  • Ratcheting under shock load. A sudden overload — collision, robot crash, drop — drives the flexspline teeth to skip over the circular-spline teeth. The unit completes the motion but loses one or more teeth of phase. The robot's encoder still reads the correct motor angle, but the output is now permanently offset. Symptoms: permanent positional error after a hard stop. Manufacturers rate a peak (single-shock) torque about three times rated torque; exceeding this even once may be enough.
  • Wave-generator bearing failure. The wave generator runs at full input speed (3000+ RPM) and supports a constantly varying elliptical load. It's the highest-cycle, highest-stress component in the unit. Failure shows up as input-side noise, increased running torque, or vibration. Once the bearing fails, the elliptical shape is no longer maintained and tooth engagement collapses — usually a unit replacement, not a rebuild.
  • Lubricant breakdown. Harmonic drives use a proprietary high-viscosity grease (HDSI's SK-1A or equivalent) that combines low-temperature flow with high-pressure film strength under the elliptical loading. Substituting general-purpose grease almost always shortens life by 50% or more. Vacuum/aerospace units use dry-film MoS₂ coatings on the flexspline teeth instead, with their own service-life limits.
  • Misalignment between motor and wave-generator. An off-axis input shaft puts side-load on the wave-generator hub and unbalances the flexspline ellipse. Symptom: cyclic torque ripple at input frequency. Fix: precision pilot-bore at the motor flange, with concentricity under 0.02 mm.
  • Reverse-driving when not designed for it. Harmonic drives are bidirectional but the wave generator is the highest-speed component; back-driving the output applies the speed multiplication to the flexspline teeth in reverse, exposing them to side loads they weren't designed for. Brakes on the input side are the textbook solution.

Frequently asked questions

Why does a harmonic drive have nearly zero backlash?

Because dozens of teeth are engaged simultaneously on opposite sides of the flexspline, not just one or two. Each tooth contributes a small fraction of the load, so manufacturing-tolerance gaps in any one tooth get averaged out. Pre-loading the wave generator further closes the residual gaps. Production harmonic drives quote backlash under 30 arc-seconds — about 0.008° — versus 5 to 30 arc-minutes for a comparable planetary gearbox.

How is the reduction ratio computed?

Ratio i = N_flex / (N_circ − N_flex) when the circular spline is fixed and the flexspline is the output. With the standard 2-tooth difference (N_circ = N_flex + 2), this simplifies to i = N_flex / 2. A 200-tooth flexspline against a 202-tooth circular spline gives 100:1. A 160-tooth pair gives 80:1. The number of teeth on the flexspline literally is twice the reduction ratio.

Why is the flexspline thin-walled?

Because it has to flex into an ellipse and back into a circle on every wave-generator revolution — twice for each rotation of the input. A thin radial wall (typically 1 to 3% of the diameter) lets it bend without exceeding fatigue limits. Too thin and the cup deflects under torque, costing torsional stiffness; too thick and the cyclic bending stress fatigues out the cup. Manufacturers spend enormous engineering effort on this single dimension.

Where do harmonic drives fail?

Three places. First, fatigue cracks in the flexspline cup wall, usually starting at the open-end rim and propagating back toward the closed end — accelerated by lubricant breakdown or contamination. Second, ratcheting: under sudden overload (collision, drop), teeth can climb out of mesh and skip, leaving the system positionally lost. Third, wave-generator bearing failure: the elliptical bearing is the most stressed component and runs at full input speed, so it sees more load cycles than anything else in the unit.

Why are harmonic drives the default in robot wrists?

Three properties stack: (1) very high single-stage ratio so a small motor and a small reducer give a stiff output, (2) flat pancake form factor so the joint is short axially, (3) zero backlash for repeatable positioning. ABB, KUKA, Fanuc, and Yaskawa industrial robots all use harmonic drives at the smaller wrist joints (joints 4-6) and often planetary or RV-style cycloidals at the heavy proximal joints (joints 1-3) where torque density matters more than backlash.

How do you choose between harmonic, planetary, and cycloidal reducers?

Harmonic for the highest single-stage ratio (30:1 to 320:1), zero backlash, and thin axial profile — robot wrists, satellite gimbals, semiconductor stages. Planetary for the highest torque density and shock tolerance, with backlash typically 5 to 15 arc-minutes and ratios up to 100:1 in three stages. Cycloidal (RV-style) for industrial robots' main joints where a combination of high torque, low backlash (under 1 arc-minute), and shock-resistance is wanted at higher cost.