We sell harmonic drives, so you'd expect us to tell you harmonic drives are always the answer. They aren't: we also build planetary servo joints, precisely because some axes are better served by one. What follows is the framework we use ourselves when an engineer sends us an application: five questions, asked in order, that eliminate the wrong families fast.
The one-paragraph version
If your positioning spec is written in arcseconds, or your load reverses and the joint must not knock, you want a strain wave gear. If you need a ratio under about 10:1, a high output speed, or the lowest cost per newton-metre, you want a planetary. If you're driving the base axis of a machine that weighs as much as a car and takes crash loads for a living, that's cycloidal territory. Everything below is the justification.
What actually differs
Typical figures for precision-grade units of comparable frame size. Use them for first-pass elimination, then work from real datasheets:
| Property | Strain wave (harmonic) | Planetary | Cycloidal |
|---|---|---|---|
| Backlash | Zero-backlash class, lost motion in arcsec | 1–15 arcmin (precision grades ~1–3′) | ~1 arcmin, preloaded |
| Single-stage ratio | 50–160:1 | 3–10:1 | 30–120:1 |
| Torsional stiffness | Moderate (the flexspline is a spring) | High | Very high |
| Shock overload | Limited (ratcheting is the hard ceiling) | Good | Excellent (~5× rated momentary) |
| Weight for a given ratio | Lightest | Middle | Heaviest |
| Efficiency at rated load | 75–90% | ~95% per stage | 80–90% |
| Cost | High | Low | High |
Question 1: what does your positioning spec say?
This is the question that decides most joints, and the gap is bigger than the units suggest. From our own catalogue: the AS-SJ1-40 harmonic servo joint holds backlash to ≤15″. Our AS-PJ-32 planetary servo joint — a good unit, honestly specified — is ≤12′. Same catalogue, same test bench: that's a 48× difference.
Put it at the end of a 1-metre arm and the abstraction disappears: 15 arcseconds of backlash is about 0.07 mm of free play at the tool. Twelve arcminutes is 3.5 mm. If your process cares about tenths of a millimetre at the end effector — dispensing, insertion, inspection, surgery — the planetary is eliminated before you reach question two. If the axis just needs to swing a camera mount to roughly the right place, paying the harmonic premium is buying precision you'll never use.
Question 2: what ratio do you actually need?
A strain wave gear can't do small ratios: the tooth-count arithmetic starts around 30:1 and our range runs 51:1 to 161:1. A planetary can't do big ones in a single stage: past ~10:1 you're stacking stages, and every stage adds backlash, length, and loss. The AS-PJ-32 uses a single 6:1 stage exactly because that axis wants output speed, not torque multiplication, a ratio no strain wave gear can offer.
Question 3: what happens on the worst day?
A crash, an e-stop from full speed, a tool jam. A planetary shrugs off transient overloads well; a cycloidal is built for them. A strain wave gear has a hard ceiling: exceed the momentary peak torque and the flexspline teeth can jump the mesh — ratcheting — which permanently degrades accuracy. It's a known, published limit, and designing to it is routine (we cover it in how harmonic drives fail), but if your duty cycle is defined by impact loads, the harmonic is the wrong family.
Question 4: how much mass can the joint carry?
In a serial arm, every gram in joint 4 is torque that joints 1–3 must lift. This is where the strain wave gear's single-stage ratio pays: our AS-SL-52 lightweight joint puts 33 N·m peak in 375 g with a magnesium housing. Matching that ratio with a multi-stage planetary of equal precision costs length and mass at every stage. This is why arms and wrists go harmonic even when budgets are tight: the mass saving cascades down the whole kinematic chain.
Question 5: force control, or position control?
If the joint must feel what it's touching — polishing, assembly with compliance, physical human interaction — reducer friction and wind-up stand between the motor current and the truth. One answer is measuring output torque directly: our force-sensing joints put a ±35 N·m sensor at 0.5% accuracy on the output side of an AS-SF-70, downstream of the reducer entirely. If force fidelity is your defining requirement, select for the sensing architecture first and the reducer second.
Why mixed architectures are normal
Look inside a current humanoid or cobot and you'll usually find both families: strain wave gears in the arms and wrists where precision and mass dominate, planetaries in legs, grippers, and wheel drives where shock, speed, and cost dominate. That's not indecision; it's the framework above applied per axis. Specify each joint on its own requirements, not on a platform-wide default.
Undecided between two candidates? Send us the load case: torque, speed, envelope, and what the axis has to do. Sizing both options and quoting is a same-working-day job.
Related: How harmonic drives fail · Torsional stiffness is a curve.