A Harmonic Drive is a strain wave gearbox that produces high reduction ratios — typically 30:1 to 320:1 — in a single compact stage with near-zero backlash. The flexspline, a thin-walled steel cup with external teeth, is the heart of the unit; it deflects elastically under an elliptical wave generator so its teeth mesh with a rigid circular spline at two opposing points. This solves the problem of getting precise, stiff reduction into a robot wrist or telescope drive without a multi-stage gear stack. You see it on every modern industrial robot — FANUC, ABB, KUKA — and on the Mars rover wheel actuators.
Harmonic Drive Interactive Calculator
Vary flexspline teeth, tooth-count difference, and input turns to see the harmonic drive reduction and slow reverse output motion.
Equation Used
The calculator uses the harmonic drive tooth-count difference. With the circular spline fixed, the flexspline turns backward by d teeth per wave-generator revolution, so the reduction magnitude is Nf/d. The article example uses Nf = 200 and d = 2, giving 100:1 and -2 flexspline teeth per input revolution.
- Circular spline is fixed.
- Flexspline output is opposite the wave generator direction.
- Tooth-count difference d is the circular spline teeth minus flexspline teeth.
Operating Principle of the Harmonic Drive
Three parts do all the work — a wave generator, a flexspline, and a circular spline. The wave generator is an elliptical steel plug fitted with a thin ball bearing that lets the outer race deform into the same ellipse. Slip that inside the flexspline — a cup-shaped steel sleeve with external gear teeth — and the flexspline takes on the elliptical shape too. The circular spline is a rigid internal ring gear with 2 more teeth than the flexspline. At the two long-axis points of the ellipse, the flexspline teeth fully engage the circular spline. At the short-axis points they sit clear. Spin the wave generator one full turn and the flexspline rotates backward by exactly 2 teeth relative to the circular spline. That's where the huge reduction comes from — you're not counting gear teeth meshing, you're counting tooth-count *difference*.
The ratio works out to -Nf / 2, where Nf is the flexspline tooth count. A 200-tooth flexspline gives 100:1 reduction in a single stage. No idlers, no compound trains, no stacked planetaries. The trade-off is that the flexspline flexes elastically every revolution of the input — typical strain at the major axis is around 0.3% — so flexspline fatigue life sets the gearbox life, usually 7,000 to 35,000 hours depending on load profile.
Get the tolerances wrong and the symptoms show up fast. If the wave generator bearing wears out and the ellipse loses its shape, you lose tooth engagement at the major axis and the unit develops audible ratcheting under torque. If the flexspline is over-strained — usually from shock-loading past the rated peak torque — you'll see tooth-tip fatigue cracks at the major axis after just a few hundred hours. Lubricant degradation is the other big killer; the proprietary grease (SK-1A or equivalent) carries the entire tooth-mesh load and once it shears down, tooth wear accelerates exponentially. Torsional stiffness and lost motion both degrade together as the unit ages — a fresh CSF-25 might show 0.5 arc-min lost motion, a worn one 3+ arc-min.
Key Components
- Wave Generator: An elliptical steel cam wrapped in a thin-section ball bearing. The bearing's outer race deforms with the cam, which lets the wave generator spin freely inside the flexspline while imposing the elliptical shape. The bearing is the consumable element — typical L10 life is 7,000–10,000 hours under rated load.
- Flexspline: A thin-walled steel cup with external teeth on the open end and a solid mounting flange on the closed end. Wall thickness is usually 0.5–1.5 mm depending on size, and tooth-form tolerance must hold ±5 µm or tooth engagement gets sloppy at the major axis. Strain at the major axis runs around 0.3% — well below yield, but high enough that fatigue eventually wins.
- Circular Spline: A rigid internal ring gear with Nf + 2 teeth. It bolts to the gearbox housing in standard configurations. Tooth profile is a modified involute ground to ±3 µm to match the flexspline at the major axis without binding at the short axis.
- Cross-Roller Output Bearing: Most modern strain wave units integrate a cross-roller bearing on the output flange to handle the moment loads from a robot arm or rotary table directly. It eliminates the need for an external bearing and keeps the package short — often under 50 mm axial length for a CSG-25 size.
Where the Harmonic Drive Is Used
Harmonic Drive units dominate any application where you need precise rotational positioning, high torque density, and zero-backlash output in a tight envelope. Robot joints are the obvious one, but the technology shows up anywhere a multi-stage planetary would be too long, too heavy, or too imprecise. The high ratio in a single stage is what sells it — a 100:1 reduction in 40 mm of axial length is hard to beat with any other gear reducer.
- Industrial Robotics: Joint reducers on FANUC M-20iA and ABB IRB 1200 6-axis arms — typically CSF-25 or CSG-32 units delivering 100:1 to 160:1 at the wrist axes.
- Aerospace: Wheel and steering actuators on the NASA Curiosity and Perseverance Mars rovers, where the zero-backlash and vacuum-compatible variant survives temperature swings from -120°C to +20°C.
- Semiconductor Manufacturing: Wafer-handling robots from Yaskawa and Brooks Automation, where particle generation must stay below ISO Class 1 cleanroom limits.
- Optical and Defense: Azimuth and elevation drives on satellite ground-station antennas and on naval gun turret stabilisers — the high torsional stiffness keeps pointing accuracy under wind and recoil loads.
- Medical Imaging: Gantry rotation drives on certain Siemens and GE CT scanner sub-assemblies, plus surgical robot joint actuators on the Intuitive Surgical da Vinci platform.
- Machine Tool: Indexing tables and rotary B-axis drives on Mazak and DMG MORI 5-axis machining centres, where lost motion below 1 arc-minute is a hard requirement for contouring accuracy.
The Formula Behind the Harmonic Drive
The single most useful equation for a Harmonic Drive is the reduction ratio. It tells you the input speed you need to hit a target output speed, and it tells you the torque multiplication you'll get on the output flange. Where the design sweet spot sits depends on the application — at low ratios near 30:1 the flexspline strain stays modest and life is long, but you give up torque density. At high ratios near 320:1 you get massive torque per kilogram, but flexspline strain is near the limit and shock-load tolerance drops. Most robot wrist designs land at 80:1 to 160:1 because that's where torsional stiffness, ratio, and life balance.
Variables
| Symbol | Meaning | Unit (SI) | Unit (Imperial) |
|---|---|---|---|
| i | Reduction ratio (negative sign indicates output rotates opposite to input when circular spline is fixed) | dimensionless | dimensionless |
| Nf | Number of teeth on the flexspline | teeth | teeth |
| Nc | Number of teeth on the circular spline (always Nf + 2 in standard units) | teeth | teeth |
| ωout | Output angular velocity at the flexspline flange | rad/s | RPM |
| ωin | Input angular velocity at the wave generator | rad/s | RPM |
Worked Example: Harmonic Drive in a 6-axis arc-welding robot wrist
You are sizing the wrist-axis reducer on a custom 6-axis arc-welding robot built around a 750 W servo motor. The motor runs nominally at 3000 RPM and you need the wrist to slew at 30 RPM nominal, with a peak torque demand of 120 Nm at the welding torch. You're choosing between a CSF-25 unit with a 200-tooth flexspline (100:1) and you want to verify the input speed and torque numbers across the realistic operating range — from a slow weave at 10 RPM output up to a 60 RPM rapid repositioning move.
Given
- Nf = 200 teeth
- Nc = 202 teeth
- ωout,nom = 30 RPM
- Tout,peak = 120 Nm
- η = 0.80 efficiency
Solution
Step 1 — calculate the reduction ratio from the tooth counts:
The negative sign tells you the output flange spins opposite the wave generator input. For sizing purposes you work with the magnitude, 100:1.
Step 2 — at nominal 30 RPM output, find the input speed the servo must deliver:
That's a perfect match to the servo's rated speed — the wrist slews cleanly at the motor's continuous operating point, and the welding seam tracks smoothly without speed ripple.
Step 3 — at the low end of the operating range, 10 RPM output for a slow weave pattern:
The servo runs at one-third of rated speed. Torque capacity is fine but cogging and encoder resolution start to matter — at 1000 RPM motor speed with a 20-bit encoder, you're still resolving sub-arc-second wrist positions, so weave quality stays clean.
Step 4 — at the high end, 60 RPM output for fast repositioning between weld passes:
This is double the servo's rated speed and above the input speed limit of the CSF-25 (typically 5600 RPM continuous, 8500 RPM intermittent). You can hit 60 RPM output for short reposition moves but not continuously — the wave generator bearing heats up rapidly above rated input speed and grease life collapses.
Step 5 — required input torque at peak load, including efficiency loss:
A 750 W servo at 3000 RPM delivers about 2.4 Nm continuous, so you have 60% torque headroom at peak welding load.
Result
The CSF-25 at 100:1 needs 3000 RPM input to deliver 30 RPM output, and the servo needs 1. 50 Nm to push 120 Nm at the torch. That's a clean match — the motor runs at its sweet spot during normal welding and has torque margin for transients. Across the range, 10 RPM weave moves drop the motor to 1000 RPM where encoder resolution and cogging behaviour start to dominate path quality, and 60 RPM rapid moves push the input to 6000 RPM, above the gearbox's continuous rating, so those moves must be short-duration only. If you measure higher-than-predicted lost motion at the torch tip — say 2 arc-minutes instead of the spec'd 0.5 — check the wave generator bearing preload first (worn bearings let the ellipse collapse under load), then the flexspline mounting bolt torque (loose flange bolts add 1+ arc-min of compliance), and finally inspect for grease shear-down which shows up as black grease colour and metallic glitter when you wipe a sample on white paper.
Harmonic Drive vs Alternatives
Harmonic Drive isn't always the right answer. Planetary reducers and cycloidal reducers cover overlapping ground, and each wins on different specs. Here's how the three stack up on the dimensions that actually drive selection decisions.
| Property | Harmonic Drive | Planetary Gearbox | Cycloidal Reducer |
|---|---|---|---|
| Single-stage ratio range | 30:1 to 320:1 | 3:1 to 10:1 | 10:1 to 200:1 |
| Lost motion / backlash | 0.5–2 arc-min (near zero) | 3–15 arc-min | 0.5–3 arc-min |
| Torsional stiffness (typical mid-size) | 1.5 × 10⁴ Nm/rad | 5 × 10⁴ Nm/rad | 3 × 10⁴ Nm/rad |
| Peak torque tolerance vs rated | 3× rated (brief) | 2.5× rated | 5× rated (shock-tolerant) |
| Efficiency | 70–85% | 92–97% | 80–90% |
| Service life under rated load | 7,000–35,000 hours | 20,000+ hours | 15,000–30,000 hours |
| Cost (mid-size, OEM volume) | $$$ — $1,200–$3,000 | $ — $200–$600 | $$ — $800–$2,000 |
| Best application fit | Robot joints, precision indexing | High-speed servo, conveyor drives | Heavy industrial robots, shock-loaded indexers |
Frequently Asked Questions About Harmonic Drive
That's almost always preload loss in the cross-roller output bearing, not the gear teeth themselves. The flexspline-circular-spline mesh is symmetric — it doesn't favour one direction. But the cross-roller bearing on the output flange develops asymmetric clearance as the rollers wear unevenly, especially if the load profile is biased (a robot arm carrying a fixed payload mostly works against gravity in one direction).
Quick check: lock the input shaft, apply a known torque to the output in both directions while measuring rotation with a dial indicator at a 100 mm radius. If forward-vs-reverse compliance differs by more than 20%, the output bearing is your problem, not the strain wave gearing.
Single 100:1 wins almost every time for a precision application. Adding a belt stage introduces 0.5–2 arc-min of backlash and 3–8 arc-min of compliance from belt stretch, which swamps the Harmonic Drive's near-zero lost motion. You've paid for the precision then thrown it away upstream of the gearbox.
The exception: if you need to physically separate the motor from the joint (long robot link, motor at the base for inertia reasons), the belt stage is unavoidable and you live with the compromise. In that case, use a steel-cord HTD belt at high preload and accept that final positioning accuracy will be belt-limited, not gearbox-limited.
60°C is at the upper edge of normal but not yet a failure. Harmonic Drive units have 70–85% efficiency, so 15–30% of input power becomes heat. A 750 W servo at 50% duty puts roughly 60–110 W of heat into the gearbox, which a CSF-25 housing can shed at around 50–65°C steady-state in still air.
The concern threshold is 80°C — above that, the proprietary grease (SK-1A or 4B No.2) starts to shear down faster than its design rate, and you'll see tooth wear accelerate. If you're hitting 75°C+, either reduce duty cycle, add a small fan, or step up to the next frame size. Don't substitute a generic EP grease — the additive package in the OEM grease is specifically formulated for the high contact stress at the major-axis tooth engagement.
The kinematic ratio is exact — 100 input revs produce exactly 1 output rev, no slip is possible because the teeth are in positive engagement. What you're seeing is torsional windup, not ratio error. Apply load to the output and the flexspline twists elastically; release load and it springs back. On a CSF-25 at rated torque you'll see roughly 1.5 arc-min of windup, which a position encoder on the motor side can't see.
Fix: put the encoder on the output flange (a true dual-encoder configuration) and close the position loop on output-side feedback. That's why high-end robots like the KUKA KR series use secondary encoders at every joint.
Mechanically yes, but it's a bad idea in most cases. Back-driving means you're now trying to spin the input shaft at 100× the output speed, and the wave generator bearing wasn't designed to operate as a high-speed output. Bearing life collapses.
More practically, the unit is non-backdrivable in the useful sense — efficiency in the back-drive direction drops to 30–50%, so a load on the output won't spin the input shaft on its own. That's why Harmonic Drive units are popular in robotics: cut motor power and the arm holds position by friction alone, no brake required for static holding.
Ratcheting on a brand-new unit usually means tooth disengagement at the major axis under load — the flexspline is deflecting radially inward away from the circular spline because something is pre-loading it incorrectly. Three causes in order of likelihood: (1) the wave generator is installed off-centre, often from a dirty mounting bore; (2) the flexspline mounting flange is bolted down unevenly, distorting the cup; (3) the output coupling is applying a side load through the flexspline cup wall.
Pull the unit apart, clean every mating surface to a clean dry condition, torque the flange bolts in a star pattern to the spec'd value (typically 35–50 Nm for an M5 grade 12.9 on a CSF-25), and verify the coupling alignment with a dial indicator before reinstalling. Ratcheting that survives proper installation means a damaged flexspline — replace it, don't try to run it in.
References & Further Reading
- Wikipedia contributors. Strain wave gearing. Wikipedia
Building or designing a mechanism like this?
Explore the precision-engineered motion control hardware used by mechanical engineers, makers, and product designers.
