Every integrated servo joint has an encoder on the motor shaft; the commutation and velocity loops can't run without one. The real architecture question is whether to add a second absolute encoder on the output flange, downstream of the reducer. It roughly doubles the feedback cost of the joint, and on some axes it's money wasted; on others it's the difference between a machine you trust and one you re-home nervously every morning.
What the motor-side encoder can't see
Through a 100:1 strain wave gear, a motor-side count is worth a hundredth of an output degree: on paper, spectacular resolution. But three things live between the motor shaft and the load, and the motor encoder is blind to all of them:
- Wind-up. The flexspline is a spring; under load the output lags the scaled motor position elastically. The motor encoder reports the commanded geometry, not the deflected one.
- Lost motion drift. As the gear wears over thousands of hours, lost motion grows (the wear clock). A motor-side-only joint has no way to notice.
- Events. After a crash or a suspected ratcheting overload, the motor-to-output relationship may have permanently shifted. With one encoder, you find out from your scrap rate.
Resolution is not accuracy
Bit-depth arithmetic first, because datasheets lean on it. One count of an absolute encoder spans 360° divided by 2bits:
| Encoder | Counts per rev | One count equals |
|---|---|---|
| 17-bit | 131,072 | 9.9″ |
| 18-bit | 262,144 | 4.9″ |
| 19-bit | 524,288 | 2.5″ |
| 23-bit | 8,388,608 | 0.15″ |
Now the honesty: those are quantisation steps, not positioning truth. A 19-bit output encoder resolves 2.5″, but the joint around it is specified for repeatability of ≤10″ or ≤20″ depending on frame. The mechanics, not the electronics, set the floor. Extra bits below that floor buy smooth velocity estimation and clean servo behaviour, not extra process accuracy. (The distinction between resolution, repeatability, and accuracy is its own article.)
What the second encoder actually buys
- Truth under load. The output encoder measures the joint after wind-up. For axes that hold position against varying torque — a nodding telescope, a loaded wrist — the control loop can close on reality.
- No homing, ever. With absolute encoders on both sides, the machine knows its pose at power-on, mid-stroke, after an e-stop. For machines embedded in production lines, eliminating the homing cycle is often the whole justification.
- Built-in condition monitoring. The difference between scaled motor position and output position is the transmission's health record: wind-up when loaded, lost motion when reversing, a step change after an overload event. Trend it and the joint reports its own ageing, with no extra sensors.
How this maps onto real hardware
In our range, the architecture is a selectable, not a philosophy: the AS-SJ2 servo joints carry dual absolute encoders — 17- or 19-bit on the motor, 19-bit on the output flange — while other frames run dual 18-bit magnetic pairs, and the AS-RA2 rotary actuators offer output resolution up to 23-bit (8,388,608 counts) for metrology-grade axes where one count needs to be smaller than the vibration floor.
The decision
Single motor-side encoder is the right call when the axis moves between taught points, the load is steady, and your accuracy budget comfortably exceeds the joint's repeatability spec: a pick-and-place shuttle, a conveyor indexer. Choose dual absolute feedback when any of these is true:
- The load torque varies while position must hold: wind-up is inside your error budget.
- Homing after every power cycle is operationally unacceptable.
- The axis is safety- or quality-critical enough that you want the transmission continuously audited.
- You need the incident record: any crash must be answerable with data, not opinion.
Unsure which side of the line your axis falls on? Send us the motion profile and the error budget. Recommending the feedback architecture is part of the sizing job, same working day.
Related: Backlash, lost motion, and stiffness · How harmonic drives fail.