A semiconductor wafer-handling arm needed a rotary axis with less than 1 arc-min of backlash and a repeatability of 5 arc-sec. The first attempt used a planetary gearbox — it measured 6 arc-min of backlash, the best the vendor could do at that ratio. The arm overshot the wafer position and the vision system flagged it every third pickup. The design switched to a harmonic drive. Backlash measured zero within the test resolution, repeatability held at 3 arc-sec, and the arm stopped missing. The harmonic drive cost 3x the planetary. For this application it was the only choice.
The mechanism
A harmonic drive has three parts. The wave generator — an elliptical cam that fits inside the second part, the flexspline — a thin-walled, flexible cup with external teeth. The third part is the circular spline, a rigid ring with internal teeth.
As the wave generator turns, it pushes the flexspline into an elliptical shape. The flexspline teeth engage the circular spline teeth at the two ends of the ellipse’s long axis. At the short axis, the teeth disengage. Because the flexspline has two fewer teeth than the circular spline, each full rotation of the wave generator advances the flexspline by two teeth relative to the circular spline. That’s the reduction. For a 100-tooth circular spline and a 98-tooth flexspline, the ratio is 50:1.
The teeth are always in contact at the load-bearing points — there’s no gap to create backlash. That’s the zero-backlash claim, and it’s real: the flexspline tooth is pushed into the circular spline tooth by the wave generator’s preload. There’s nowhere for the teeth to separate.
The numbers that matter
| Parameter | Harmonic drive | Planetary (2-stage) |
|---|---|---|
| Backlash | 0 (within measurement) | 5-10 arc-min |
| Repeatability | 2-5 arc-sec | 15-30 arc-sec |
| Ratio range | 30:1 to 160:1 (single stage) | 10:1 to 100:1 (multi-stage) |
| Efficiency | 70-85% | 90-95% |
| Torque density | High (compact) | Moderate |
| Backdrivability | Poor to moderate | Good |
| Cost per unit | 3-5x planetary | Baseline |
| Lifespan at rated torque | 5000-20000 hrs | 20000+ hrs |
The efficiency number catches people. A harmonic drive at 80% efficiency wastes 20% of the input as heat. For a 100 W servo, that’s 20 W of heat in the drive. The flexspline runs hot. This matters in two ways: the drive needs cooling in continuous high-torque duty, and the heat shortens the flexspline’s fatigue life. The wafer arm ran at 30% duty cycle — the drive stayed cool enough.
The failure mode: flexspline fatigue
The flexspline is a thin cup that flexes continuously. It has a finite fatigue life. The manufacturer rates it in cycles at rated torque. Run it at rated torque continuously and it might last 10,000 hours. Run it at 50% torque and it lasts 5x longer. The failure mode is a crack in the flexspline cup — it fails suddenly, without warning noise. For a safety-critical axis, the harmonic drive should be sized at 2-3x the actual load, and the runtime tracked.
The sizing rule for harmonic drives: pick the size where the continuous torque rating is at least 2x the actual continuous torque. The peak torque rating should be at least 1.5x the actual peak. The momentary overload rating (which can be 2.5x continuous) covers startup and crash conditions. The wafer arm’s axis needed 3 Nm continuous. The 50:1 harmonic drive in size 25 (rated 25 Nm… actually, size 20 rated about 18 Nm continuous) gave a 5x margin. Overkill for the cost, but the arm had zero tolerance for a flexspline crack mid-process.
Ratio selection and the motor side
Harmonic drives come in ratios from 30:1 to 160:1 in a single stage. The ratio determines the reflected inertia at the motor — higher ratio, higher reflected inertia. The motor sizing needs the reflected inertia to be within the servo’s inertia ratio range. The wafer arm used a 50:1 ratio with a 200 W servo — the reflected load inertia was 8x the motor inertia, which the servo handled.
The other ratio consideration is speed. A harmonic drive’s max input speed is 3000-7000 RPM depending on size. The output speed at 50:1 with a 3000 RPM motor is 60 RPM. For a fast rotary axis, that’s the limit. If the application needs faster output, either a lower ratio (which reduces the precision advantage — backlash stays zero but the resolution drops) or a different mechanism entirely.
Backdrivability and the brake question
Harmonic drives have high static friction — they resist being driven backward. That’s a feature for holding a vertical load without a brake. But it’s also a limitation: a harmonic drive can’t be backdriven for manual positioning or for load-sensing applications. If the axis needs to be moveable by hand during setup, a harmonic drive fights you. The wafer arm didn’t need manual positioning. The vertical axis of a pick-and-place that uses a harmonic drive needs a brake anyway — the drive’s friction holds the load, but a sudden power loss with a heavy load at the end of a stroke needs a positive brake.
The maintenance reality
Harmonic drives need grease. The grease is factory-filled and rated for the drive’s life. The drive should be re-greased at the manufacturer’s interval (typically 20,000 hours for the small sizes, or when the grease shows breakdown). The flexspline’s fatigue life is the real limit — track the runtime hours like you track a bearing. When the drive approaches its rated life at the actual duty cycle, replace it. It fails suddenly. A $300 drive that fails mid-process costs $30,000 in downtime on a semiconductor line.
Harmonic drives earn their cost when backlash is the spec, not the hope. The wafer arm couldn’t hit the wafer with 6 arc-min of planetary backlash. The harmonic drive measured zero, held 3 arc-sec, and ran for years. Pay for it when you need it, size it at 2x continuous torque, track the flexspline hours, and remember the 20% efficiency loss is the price of zero backlash.
