Single-sourcing a precision reducer was a reasonable position for a long time. It is a harder one to defend now. The 2025 US tariffs on precision assemblies pushed cost through the reducer supply chain, component shortages have been pushing robot deployments in Europe and North America out by a fifth to a quarter, and several national programmes are actively funding domestic reducer capacity. Buyers are looking for a second source. Engineering is being asked what it will cost.
The honest answer is that it depends entirely on which properties differ, and the difference between a cheap qualification and an expensive one is not the one most people assume.
“Drop-in” is not one property
A reducer is an interface in four directions at once: mechanical, kinematic, electrical and environmental. A part can be identical in three and disqualified by the fourth. Treat them separately or you will end up re-testing things that never changed.
What has to match exactly
The mechanical interface. Bolt circle, pilot diameter, output flange pattern, shaft or bore fit, overall length, and the through-bore if you route cable. These are pass or fail. There is no engineering judgement available here, and no amount of margin elsewhere compensates for a bracket that does not bolt up.
The ratio. Not “close to”. Your controller’s counts-per-output-revolution, your kinematic model, and every taught position depend on it. 101:1 and 100:1 are not interchangeable, and the error is cumulative rather than a fixed offset.
The direction of rotation. A strain wave gear reverses the output relative to the input in the standard configuration. If a candidate does not, you have a sign flip in the control loop and a safety case to revisit.
What has to match by class, not by digit
These are the ones worth arguing about, because a small difference here is usually free and a large one usually is not.
| Property | Acceptable difference | Why |
|---|---|---|
| Rated torque | Equal or higher | Sets the continuous duty the joint can hold |
| Peak torque | Equal or higher | Covers acceleration and stop transients |
| Backlash / repeatability | Same grade or tighter | Directly caps positioning performance |
| Torsional stiffness | Within ~10% | Moves the joint resonance and your loop gains |
| Reflected inertia | Within ~10% | Changes the motor’s effective load and tuning |
| Max. input speed | Equal or higher | Caps output speed at your chosen ratio |
A stiffer or lower-inertia part is not automatically safe, either, because stiffness sets where the joint resonates. Both shift the resonant frequency of the joint, and a controller tuned with a notch filter at the old frequency will be filtering the wrong thing. If stiffness or inertia moves more than about ten percent, plan on re-tuning. That is a day, not a programme.
What can differ freely
Mass, efficiency, no-load running torque, noise, housing material and grease specification can all differ without touching your control loop, provided the system-level budgets still close. Mass matters on a moving axis and matters enormously on an arm where every gram cascades; it does not matter on a fixed indexing table. Efficiency changes your thermal picture and your battery, not your accuracy.
What forces a real re-qualification
Three things, and only three, reliably turn a component swap into a project:
- Anything that changes the safety case. A different emergency-stop torque rating, a different brake arrangement, or a change in back-driving behaviour means the risk assessment is reopened. This is the expensive one.
- Anything that changes the control loop enough to need re-tuning across the whole workspace. Large stiffness, inertia or backlash changes qualify. Small ones do not.
- Anything that changes a regulated claim. If the machine ships with a stated accuracy, life or duty figure, and the new part is characterised on a different basis, the claim has to be re-established rather than inherited.
Where the datasheet stops
A published specification will get you to a shortlist and no further. Ours publishes rated and peak torque, ratios, backlash or repeatability, protection rating, and stiffness where the series is characterised on it — enough to rule parts out quickly, which is most of the value. It does not publish bolt patterns, bore diameters or reflected inertia, because those are interface drawings rather than performance figures, and a table is the wrong place to read them.
So the sequence that works is: filter on the published numbers, then ask for the drawing, then ask what basis the life and accuracy figures were established on. A supplier who cannot answer the third question quickly is telling you something useful.
If you are qualifying a second source against an existing joint, send us the part number you are matching and the properties above that are fixed for you. We will tell you which of ours meet them and, more usefully, which do not.
Related: Your rated life is a coin flip · Integrated joint, or build your own?.