The Encoder Is the Loop
The Encoder Is the Loop: Absolute vs Incremental, and What Resolution Actually Buys
An encoder is the sensor that tells the controller where the shaft is, that is the whole job. Incremental counts pulses and forgets everything on power loss; absolute reads a unique code for every angle and knows where it is the instant power returns. This is what resolution buys, why it is not accuracy, why velocity comes from differentiating position and gets noisy at low speed, and what Example Labs' bench arm needs on each of its four joints.
A servo without an encoder is a person walking with their eyes closed, counting their own steps. It works, until it doesn't. One step missed, and every step after it is wrong, and nothing in the system will tell you.
An encoder is the sensor on the motor shaft that reports exactly where the shaft is. A motor with an encoder knows its position. A motor without one is guessing. And the choice the article is about, incremental or absolute, is not a choice about resolution. It is a choice about what happens when the power goes out, what happens when a pulse is missed, and what the loop can actually see at the speed you need it to run.
The words, in plain English
Encoder. A sensor that turns a shaft's rotation into an electrical signal the controller can read. Rotary encoders report a shaft's angle to a motion controller. That is the whole job, and everything fancy you do with a motor (torque control, smooth velocity profiles, holding a position to a few arc-seconds, commutating a brushless motor) rides on knowing the angle of the rotor and the load.
Incremental encoder. An encoder that produces a train of pulses as the shaft rotates, and reports change rather than position. The controller counts the pulses to infer position relative to a starting point. It does not report an absolute position, it says only that the shaft moved, by how much, and in which direction.
Absolute encoder. An encoder that assigns a unique digital code to every shaft angle, so the controller knows exactly where the axis is the moment power is applied. Power it up mid-travel and the first reading is already correct.
PPR (pulses per revolution). The number of pulses one channel produces in one full turn, the number of lines on the disc. For incremental encoders, typically 100–10,000 PPR.
CPR (counts per revolution). The number of counts the controller gets after decoding. PPR and CPR are not the same number, and a datasheet that already specifies decoded counts must not be multiplied by four again, that introduces a fourfold scaling error.
Quadrature. Two square-wave channels, A and B, offset by exactly one quarter of a cycle, 90 electrical degrees. The phase order, which channel leads, reveals the direction of motion, and counting the edges reveals position. A third channel, Z (also called index, reference, or marker), emits a single pulse per mechanical revolution and provides the fixed landmark the controller needs to establish a datum.
Gray code. A code in which adjacent positions differ by exactly one bit, so a sensor edge timing skew can delay a reading but cannot produce a wildly wrong value. The classical absolute disc uses concentric code tracks patterned in Gray code. Modern devices generally use a single finely patterned track that the readhead interpolates, with the resulting position sent out over a serial interface rather than as parallel bits.
Single-turn absolute. An encoder that gives a unique code within one revolution and starts over on the next, fine for a rotary table, useless on a ten-turn lead screw.
Multi-turn absolute. An encoder that adds a revolution counter, so the turn count survives power loss too. The counter is implemented either as a geared mechanical stage that is genuinely power-independent, or as an electronic counter kept alive by a battery or an energy-harvesting element (a Wiegand wire). A 12-bit turn counter covers 4,096 revolutions; a 16-bit counter covers 65,536.
Resolution. The smallest position increment the encoder system can report, how fine the scale is.
Accuracy. How far the reading is from the true value, over one full revolution, how far the holes are from the centre of the target. It comes from many places: position errors of the lines on the disc, mounting eccentricity of the disc, bearing runout, errors in the interpolation circuit.
Repeatability. How consistent the readings are when the same position is reached again, how tightly the holes are grouped. It is affected mainly by noise, mechanical backlash, and short-term temperature changes. Repeatability is often much better than accuracy, because systematic errors such as eccentricity and line-position error are the same every time and repeat unchanged when the same position is reached, and that also brings a benefit: systematic errors can be measured and compensated.
Interpolation. The process of dividing the space between two scale lines into finer steps, using the sine and cosine signals from the readhead (the angle within a scale pitch from arctan(sin/cos)). Interpolation gives the angle within a scale pitch, and it is also where interpolation error enters the accuracy budget.
Resolver. An analog, rugged rotary position sensor, a small transformer-like device whose sine and cosine outputs are converted to an angle by resolver-to-digital electronics. Rugged to 200 °C and above, excellent in contamination and stray fields, but in practice roughly equivalent to a 1,000 PPR encoder once the converter's resolution is counted.
SinCos. A sine/cosine analog interface, the output of a resolver, or of an encoder with sine/cosine tracks. The safety value: since sin² + cos² = 1, the connected device can calculate whether the signal is genuine, uninterrupted, and reliable, and can detect a wire break, a key limitation of incremental square-wave outputs.
Commutation. The process of energising the right motor windings at the right rotor angle. A brushless motor commutated with a coarse or noisy angle wastes current as torque ripple and heat.
FOC (field-oriented control). The control method from Article #20 that makes torque proportional to the q-axis current, and that needs the rotor angle, from the encoder, to do it.
The two families: counting, and reading
An incremental encoder does not know where it is. It counts. Each line on the disc that passes the sensor produces a pulse; the controller counts the pulses and infers position relative to a starting point. The count is lost when power is removed, the controller has no way of knowing where the shaft was when the system shut down. To recover position, the machine must perform a homing or reference run, moving the axis until it finds a known reference mark or limit switch.
On a small servo axis, homing takes a second and nobody notices. On a large gantry, a robot arm with restricted reach, or a machine that stopped mid-cut, homing can be slow, awkward, or physically unsafe, and that is where the case for absolute becomes financial rather than technical.
The other consequence of counting is that errors persist. If electrical noise injects a spurious count, or a fast move outstrips the counter input, the controller does not find out. The reported position simply drifts from reality. An incremental encoder is very sensitive to electromagnetic interference: run the encoder cable next to the VFD power cable without proper shielding, and the controller will begin to pick up false counts.
An absolute encoder does not count. It reads. Each position corresponds to a specific binary or Gray-code pattern, so the controller always knows the exact shaft position, even immediately after power is restored.
No homing run is required, because the position information is inherent to the encoder's mechanical and optical design rather than accumulated in a counter. And if the shaft is turned by hand while the power is off, 90 degrees, say, the absolute encoder will immediately show the new correct position when power returns. No homing, no reference move, no procedure.
The serial interface adds a second advantage: an absolute encoder transmits data in packets. If a packet is corrupted, the controller generates an error, but it does not add an extra metre to the coordinate. The failure is visible. The incremental encoder's failure is silent.
The trade is cost and complexity. Incremental optical encoders run roughly $5–30; magnetic encoders $3–20; absolute encoders $50–500, with proprietary protocols adding complexity. And the absolute encoder's multi-turn mechanism is a second decision: if the working range is under one revolution, a single-turn unit is complete; if the axis travels multiple turns, the turn count must survive power loss too, and the geared, battery, or Wiegand mechanism is the decision.
Resolution is not accuracy
This is the single most common encoder mistake in robotics, and it is worth stating plainly: resolution is not accuracy, and confusing the two is the single most common encoder mistake in robotics.
Resolution is how fine the grid is. Accuracy is how far the holes are from the centre. Repeatability is how tightly the holes are grouped. A fine grid does not mean the shots are accurate.
The numbers, checked:
- A 14-bit magnetic encoder reports 16,384 counts per turn, about 79 arc-seconds per count, yet its real angular accuracy might be ±0.3° to ±0.5° (1,000 to 1,800 arc-seconds) once nonlinearity and mounting eccentricity are included. The counts are precise; the angle is not. The accuracy is 14–23 times worse than the resolution.
- On a 17-bit encoder, the scale of the reading is 9.9 arc-seconds, but if the disc is mounted 3 µm off centre at a 20 mm reading radius, the reading can be as much as 31 arc-seconds away from the true value, more than 3 counts. Eccentricity is the most typical item: a disc eccentricity e at a reading radius r produces a once-per-revolution sinusoidal error of amplitude Δθ ≈ e ÷ r.
- Precision is a system issue, and it can be dominated by mechanical errors: eccentricity, backlash, straightness, and flatness. Eccentricity and gearbox backlash are the most significant contributors to mechanical errors in rotary applications, and the encoder measures the motor, not the load, so the gearbox sits between the measurement and the thing being positioned.
The practical rule, from the guides: spec the encoder on the number that matches your control need. Resolution for smooth velocity and low quantization noise. Accuracy for absolute pointing and gear-train compensation. Repeatability for return-to-home. And remember that systematic errors can be measured and compensated, if a higher-accuracy reference is used to measure the error curve over one revolution and the curve is stored in the controller for correction, accuracy can be improved substantially, because the error repeats.
The velocity problem: differentiation amplifies noise
Velocity usually comes from differentiating position, and differentiation amplifies quantization noise brutally. A uniformly distributed quantization error has variance q²/12, the standard result from Bennett's 1948 quantization analysis, the same q²/12 that gives an ADC its 6.02 dB per bit. Velocity noise scales as 1/(N·T): it falls linearly with resolution and linearly with the differencing interval.
The worked example, from the guides: a 1,000 PPR encoder (4,000 quadrature counts per revolution) running at 10 rpm produces 667 counts per second. At a 4 kHz control loop rate, that is 0.167 counts per loop cycle. Because counts are integer values, the actual count per loop cycle alternates between 0 and 1, and the velocity computed from that alternates between 0 rpm and 60 rpm per count at the 4 kHz rate. Extreme low-speed velocity noise, from a sensor that is working perfectly.
Three methods exist, and each has a different noise and latency character:
- Position differencing (fixed time window). Record position at time T₁ and time T₂, velocity = (P₂ − P₁) / ΔT. Simple, works well at medium and high speeds where multiple counts occur per measurement interval. At low speed, count noise dominates, and a longer window reduces the noise but adds latency.
- Period measurement (fixed count window). Measure the time between N consecutive encoder counts; velocity = N counts ÷ measured time. Excellent low-speed resolution, because the time measurement can be very precise (a hardware timer with MHz resolution), but at low speed the method must wait for N counts to arrive, introducing a measurement latency proportional to 1/speed, and at high speed the interval becomes very small and sensitive to timestamp jitter.
- Hybrid. Frequency measurement at low speeds, differencing at high speeds, whichever is more accurate at the speed the axis is actually running.
And the rule of thumb, from the guides: for smooth low-speed velocity control, you want at least 12-bit (4,096 CPR) resolution at the point you are controlling, and 17-bit+ (131,072 CPR) for direct-drive joints that must crawl smoothly. On a good drive, a state observer is doing more for low-speed smoothness than the encoder's headline bit count, but the observer cannot invent counts the encoder never produced.
The frequency ceiling: the number that caps the PPR
The output frequency of an incremental encoder is f = PPR × rpm ÷ 60, and it must be considered from two points of view: the maximum allowed output frequency of the encoder, and the input frequency of the pulse inputs of the controller or encoder card. The resolution of the encoder must not exceed the input frequency.
The numbers, checked: the maximum possible pulse frequency is usually 300 kHz for standard types and up to 800 kHz for high-resolution encoders (Kübler), and one drive manufacturer sets the maximum output frequency of a signal track at 1 MHz, with the guidance to select resolution so that the output frequency is almost 1 MHz at maximum machine velocity, and to keep 800 kHz (20% reserve) at maximum velocity. Standard encoders run 1–5,000 PPR; high-resolution encoders run up to 36,000 PPR.
The pipeline's figures check out exactly: a 2,048 PPR encoder at 3,000 rpm produces 102.4 kHz per channel (2,048 × 3,000 ÷ 60), and a 5,000 PPR encoder at 3,000 rpm produces 250 kHz, which the drive input must accept. And at 4× decoding, the decoder sees four times as many edges: 410 kHz for the 2,048 PPR case, and 1 MHz for the 5,000 PPR case, at the 1 MHz limit, with the 800 kHz reserve guidance breached.
The first limit a fast axis hits is the encoder's electrical frequency limit; then the mechanical limit (the disc's maximum rpm, typically 6,000 rpm for the standard types).
The sensing families: what survives your environment
The encoder is one citizen of the input landscape, and the sensing technology is chosen by environment, not by catalogue resolution:
| Technology | Typical accuracy | Max resolution | Contamination | Stray-field immunity | Relative cost |
|---|---|---|---|---|---|
| Optical | ±1 to ±20 arc-sec | 28 to 32 bit | Poor (sealing helps) | Excellent | High |
| Magnetic (on-axis) | ±0.1° to ±0.5° | 12 to 17 bit | Excellent | Poor | Low |
| Capacitive | ±0.1° to ±0.2° | 12 to 14 bit | Good (not humidity) | Excellent | Medium |
| Inductive | ±arc-min to ±0.05° | 18 to 22 bit | Excellent | Excellent | Medium-High |
| Resolver | ±5 to ±20 arc-min | RDC-set (10 to 16 bit) | Excellent | Excellent | Medium |
Optical leads on accuracy, and its weakness is the optical path: a glass disc that can shatter, electronics susceptible to heat and vibration, and an optical path that is sensitive to contamination if the sealing is inadequate. Magnetic is cheap, dust-immune, and contactless, with no optics to fail, and its weakness is the motor itself: a magnetic encoder must reject the motor's stray field and current-related disturbance, the magnet must sit within 0.5–1 mm of the chip and be aligned to about 1°, and a glued magnet can slip after thermal cycling.
The resolver is the rugged analog veteran, no optics, no electronics on the rotating part, rugged to 200 °C and above, and its weakness is the converter: the resolver-to-digital electronics set the resolution, and only a handful of drive manufacturers support resolver feedback, which is why digital encoders are squeezing them out.
The opinion in the guides, worth repeating: if you are building a robot arm and your knee-jerk is "magnetic, it's cheap and rugged", seriously look at inductive first. You keep the ruggedness, lose the stray-field headache (your motor is a stray field), and gain a half-decimal-place of accuracy.
Interfaces: the wire is part of the loop
The interface choice is made with the architecture in view:
- A/B/Z (RS-422). The incremental standard: differential pairs (A/Ā, B/B̄) for noise immunity on long cables, plus the Z index. Simple, universal, and the only interface an incremental encoder needs, but it carries no diagnostics, and a missed pulse is silent.
- SSI (Synchronous Serial Interface). The classic absolute serial link that any PLC input card reads. Point-to-point, clock up to 2 MHz, position data left-aligned. Simple and universal, but one-directional: the drive cannot ask the encoder anything.
- BiSS C. The servo-class bidirectional protocol, the newest absolute standard, built into modern servo motors and direct-drive motors, with clock frequencies up to 10 MHz and update rates into the tens of microseconds for closed servo loops. Lower-priced than EnDat, and the multi-turn variants need a battery inline (without it, single-turn absolute).
- EnDat 2.2 (Heidenhain). The other servo-class bidirectional protocol, with an 18–19 bit resolution and 4,096-turn multi-turn variants, and an analog commutation track for backwards compatibility on some models.
- HIPERFACE / HIPERFACE DSL (SICK/Stegmann). The third servo-class option, with DSL offering a single-cable design carrying both power and feedback.
- PROFINET / EtherCAT. The industrial network profiles, where position rides the fieldbus itself as process data, chosen when the encoder is one citizen of a networked machine, not a point-to-point drive connection.
The latency point: the classic SSI link that any PLC reads is fine for a PLC. The servo-class links, BiSS, EnDat, HIPERFACE, push update rates into the tens of microseconds, which is what a closed servo loop running at 10–40 kHz needs.
Functional safety: when the encoder is the safety device
When the encoder's reading is what keeps a person safe, it is not a feedback device. It is a safety device, and it must comply with the functional safety standards:
- IEC 61508, the core functional safety standard, applied to all critical safety electrical, electronic, and programmable systems; it defines the Safety Integrity Level (SIL) framework.
- IEC 61800-5-2, the standard for safe motor feedback; it defines the safety-rated drive functions (Safe Stop 1 and 2, Safe Operating Stop, Safe Limited Speed, Safe Direction, Safely Limited Position) and the requirements for adjustable-speed electrical power drive systems.
- ISO 13849-1, the safety of machinery and safety-related parts of control systems; it defines the Performance Level (PL) framework.
- IEC 62061, the safety of machinery and functional safety, based on EN 61508.
The certified parts exist: Renishaw's RESOLUTE FS (BiSS Safety absolute) and TONiC FS (incremental) are certified to ISO 13849 Category 3 PLd, IEC 61508 SIL2, and IEC 61800-5-2 SIL2, and support the safety sub-functions SS1, SS2, SOS, SLS, SDI, SLP, and SSM. Pepperl+Fuchs' RVS58S is the first ISO 13849-certified SinCos encoder specifically designed for use in safety-aligned systems up to SIL3 (IEC 61508, IEC 62061), performance level "e" (IEC 13849), and is suitable as a motor feedback system for safe drives in accordance with IEC 61800-5-2. And the SinCos interface is what makes the safety case work: since sin² + cos² = 1, the connected safety device can calculate whether the signal is genuine, uninterrupted, and reliable, and can detect a wire break, which is the key limitation of incremental square-wave outputs.
The certification comes with a condition, stated in the datasheets: the system must be installed and operated in accordance with the instructions in the installation guide. Failure to follow the correct use instructions, and failure to heed the limitations, may result in SIL2 and/or PLd not being achieved, and will invalidate the functional safety certification.
The sizing rule
Define the required position resolution at the load, convert to counts per revolution, and apply a 4–5× margin so control and filtering are not starved of counts. Then check the three ceilings: the output frequency at maximum speed (f = PPR × rpm ÷ 60, against the encoder's maximum and the drive input's maximum), the velocity noise at minimum speed (the counts per control-loop cycle at the creep speed), and the accuracy budget (the mechanical errors, eccentricity, backlash, wind-up, that sit between the encoder and the load).
The general guidance, from the pipeline and the guides: general servo, incremental 1,000–2,500 PPR or absolute 17-bit; high-precision CNC, absolute 20-bit or above. And the question that decides the family before any of the numbers: is homing acceptable? If the machine can easily reach the limit switch when powered on, go with incremental. If not, if there is a load on the shaft, if the axis stopped mid-cut, if the position is safety-critical, only an absolute encoder is suitable.
the bench arm's four joints, sized on paper
Example Labs is a made-up company we use for worked examples.
Example Labs' bench arm, from Articles #20–#23: two links, 0.425 m each, 4 kg and 3 kg, a 5 kg payload at 0.85 m reach. The shoulder needs 126.8 N·m with margin, through a 100:1 planetary at 93% efficiency, so the motor needs 1.36 N·m at 1,500 RPM, and the joint turns at 15 RPM (90° in 1 s). The elbow needs 40.7 N·m, the wrist 11.0 N·m. Every number below was run in Python before writing; the encoder choices are the article's own, built from the verified figures above.
Step 1: Resolution at the load, with the margin
The requirement: hold the endpoint to 0.1 mm at the 0.85 m reach. The joint angle that moves the endpoint 0.1 mm is 0.1 mm ÷ 850 mm = 0.000118 radians = 0.00674° (24.3 arc-seconds). Through the 100:1 gearbox, that is 0.674° at the motor, so the motor encoder needs 534 counts per revolution minimum, and with the 4–5× margin, 2,136–2,670 counts.
- A 1,024 PPR incremental at 4× decoding gives 4,096 counts, passes the 4× margin and the 5× margin. It is enough, on resolution.
- A 512 PPR incremental at 4× gives 2,048 counts, fails the 4× margin. It is not enough, and the failure is silent: the loop would simply have fewer counts than the control and filtering need.
For scale: the smart servo from Article #20 positions to ±0.088°, which at 0.85 m is 1.3 mm at the endpoint, and a 0.05° error at the joint is 0.74 mm. The 0.1 mm requirement is tight, and it is the requirement that sizes the encoder.
Step 2: Resolution is not accuracy, on this arm
A 17-bit absolute encoder gives 131,072 positions per turn, 0.00275° per step, 9.9 arc-seconds, and through the 100:1 gearbox that is 0.099 arc-seconds per count at the joint. It looks absurdly fine. But the accuracy is not the resolution:
- A 14-bit magnetic encoder reports 16,384 counts per turn (79 arc-seconds per count) yet can be off by ±0.3° to ±0.5°, 1,080 to 1,800 arc-seconds, 14–23 counts of error. The counts are precise; the angle is not.
- A 3 µm mounting eccentricity at a 20 mm reading radius produces a 31 arc-second error, more than 3 counts on the 17-bit encoder. The reading can be 3 counts away from the true value while the resolution claims 9.9 arc-seconds.
And the error that dominates them all is not in the encoder at all: the harmonic gearbox from Article #21 winds up 10.8 mm at the 0.85 m reach under the 126.8 N·m shoulder load, 43 times its own backlash. The encoder measures the motor. The load is after the gearbox. No encoder, however fine, can measure a wind-up that happens between the sensor and the load. The accuracy budget must include the gearbox, and on this arm, the gearbox is the budget.
Step 3: The velocity loop, at creep speed
The shoulder motor at 1,500 RPM with a 1,024 PPR incremental at 4× decoding produces 102,400 counts per second, 25.6 counts per 4 kHz velocity-loop cycle. At full speed, the velocity estimate is smooth.
Now the creep: the joint moves at 0.1° per second for fine positioning, so the motor turns at 10° per second, 114 counts per second, or 0.028 counts per 4 kHz cycle. The velocity estimate alternates between 0 and one count, and one count per cycle is 351.6° per second at the motor, 3.52° per second at the joint, 35 times the intended 0.1° per second. The controller would be reacting to a measurement staircase 35 times larger than the motion it is trying to control.
With a 17-bit absolute encoder (131,072 counts): one count per 4 kHz cycle is 10.99° per second at the motor, 0.110° per second at the joint, the same order as the intended 0.1° per second, and the creep produces 0.91 counts per cycle. The 17-bit encoder makes the creep controllable. The 1,024 PPR encoder does not, at the velocity loop's raw differencing. This is what the resolution is bought for: not the position, but the velocity.
Step 4: The frequency ceiling, at full speed
At the shoulder's 1,500 RPM, a 2,048 PPR encoder produces 51.2 kHz per channel, comfortably inside the 300 kHz standard limit. But the same encoder at 3,000 RPM produces 102.4 kHz per channel, 410 kHz of edges at 4× decoding, and a 5,000 PPR encoder at 3,000 RPM produces 250 kHz per channel, 1 MHz of edges, at the 1 MHz limit, with the 800 kHz reserve guidance breached. And a 5,000 PPR encoder at 6,000 RPM produces 500 kHz per channel, 2 MHz of edges, past every limit in the guides, and past the disc's 6,000 RPM mechanical limit as well.
The PPR that the velocity loop wants at low speed is the PPR that the frequency ceiling forbids at high speed. That is the trade, and it is why the absolute encoder, whose serial interface does not have a per-channel pulse frequency, wins on fast axes.
Step 5: The homing cost, on a loaded arm
The incremental encoder loses position on power loss, and the bench arm must home. Homing the shoulder through 90° at 15 RPM takes a second of motion, but on a loaded arm with a 5 kg payload, the homing move is a real motion with real risk: under position control, the arm's peak contact force in Article #22's simulation was 1,159 N. The homing move is not free. It costs a second, a risk, and a procedure, and on a machine that stopped mid-cut, it may cost the workpiece.
The absolute encoder costs more, $50–500 against $5–30 for an incremental optical encoder and $3–20 for a magnetic one, but it removes the homing move entirely. And on this arm, the multi-turn question answers itself: the shoulder and elbow travel less than one revolution, so single-turn absolute is enough for all four joints. Multi-turn is for lead screws and winches, not for joints.
Step 6: The selection, per joint
- Shoulder: 17-bit absolute, BiSS C. The creep needs the resolution, the loaded arm cannot home safely, and the safety case (Article #22's force limits) wants a position the drive can trust at power-up. The cost is justified by the joint that carries the most torque.
- Elbow: 17-bit absolute, BiSS C. Same reasons, smaller numbers, and the same interface, so the drive sees one encoder family.
- Wrist: 1,024 PPR incremental, RS-422 A/B/Z. From Article #20, the wrist is a stepper with a 10:1 planetary, and the stepper's own open-loop step is 1.8°, the 4,096 counts from 4× decoding are more than the wrist needs, the load is small, and homing the wrist is a small, safe motion.
- Gripper: 12-bit magnetic absolute, SPI. Cheap ($3–20), dust-immune, no optics to fail, and the gripper's requirement is "closed or open", not 0.1 mm. The 12-bit step is 0.088°, which is finer than the gripper will ever need, and the magnetic encoder's ±0.1–0.5° accuracy does not matter when the requirement is a state, not a position.
The choice is per joint, not per robot, the same conclusion as Article #20's motor selection, for the same reason. The encoder is bought for the loop it closes, and each joint closes a different loop.
The decision test
Five questions, in order. If you cannot answer the first, the rest do not matter.
- Is homing acceptable? If the machine can easily reach the limit switch when powered on, go with incremental. If there is a load on the shaft, if the axis stopped mid-cut, if the position is safety-critical, only an absolute encoder is suitable.
- What is the required resolution at the load, with the 4–5× margin? Define it at the load, convert to counts per revolution, apply the margin. A 512 PPR encoder that fails the margin fails silently.
- What is the velocity noise at the creep speed? Count the counts per velocity-loop cycle at the slowest speed the axis must run smoothly. If it is under one count per cycle, the velocity estimate is a staircase, and no amount of PID tuning will fix it.
- What is the output frequency at maximum speed? f = PPR × rpm ÷ 60, per channel, and four times as many edges at 4× decoding. Check it against the encoder's maximum and the drive input's maximum, with the 20% reserve.
- What is the accuracy budget, including the mechanics? The encoder's resolution is not its accuracy, and neither is the system's accuracy, the gearbox's backlash and wind-up, the mounting eccentricity, and the bearing runout all sit between the sensor and the load. If the gearbox winds up 10.8 mm, a 0.099 arc-second count is measuring the wrong thing precisely.
Objections
"But the absolute encoder costs ten times the incremental." Ten times a $5–30 part is $50–500, and the homing move it removes costs a second, a risk, and a procedure on every power cycle. On a machine that stops mid-cut, the homing move can cost the workpiece. The encoder is not the expensive part of the axis. The motor, the gearbox, and the drive are. Buying the cheapest sensor for the most expensive joint is the wrong place to save money.
"But we already have 17-bit encoders, so our accuracy is 0.0027°." No. Your resolution is 0.0027°. Your accuracy is the sum of the encoder's nonlinearity, the mounting eccentricity (3 µm at 20 mm is 31 arc-seconds, more than 3 counts), the bearing runout, and the gearbox's wind-up, and on this arm, the wind-up alone is 10.8 mm at the endpoint. The 17-bit encoder is bought for the velocity loop, not for the position. Say which one you are buying it for.
"But the magnetic encoder is cheap and rugged, so we use it everywhere." It is cheap and rugged, and it must reject the motor's stray field, which is the field of the motor it is mounted on. If your knee-jerk is magnetic, look at inductive first: you keep the ruggedness, lose the stray-field headache, and gain a half-decimal-place of accuracy.
"But we can just filter the velocity noise." You can, and a good drive's state observer does more for low-speed smoothness than the encoder's headline bit count. But the observer cannot invent counts the encoder never produced, and the filter adds latency to the loop that the observer is trying to make faster. The fix is resolution at the point you are controlling, not filtering after the fact.
Sources
- Industrial PLC Systems, "Incremental vs Absolute Encoders: What Happens After Power Loss"
What we used
An absolute encoder remembers its position even after power loss; an incremental encoder tracks movement from a starting point, is useful for speed and direction, but forgets everything the moment the power cuts; the incremental count is lost when power is removed and the machine must perform a homing or reference run; an absolute encoder reports a unique position word for every distinct shaft angle, each position corresponding to a specific binary or Gray-code pattern, so the controller always knows the exact shaft position even immediately after power is restored, with no homing run required; if the machine must resume production immediately after a power dip, or if the axis position is safety-critical, an absolute encoder eliminates the need for reference moves
- DifferenceBetweenZ, "Absolute vs Incremental Encoders: Differences"
What we used
An absolute encoder gives every position its own unique code, so the controller knows exactly where the axis is the moment power is applied; an incremental encoder emits pulses as it moves, reports change rather than position, and must be homed against a reference mark after every power cycle; older absolute designs use parallel tracks of concentric rings, typically patterned in Gray code so that only one bit changes between adjacent positions and a misread boundary cannot produce a wildly wrong value; modern devices generally use a single finely patterned track that the readhead interpolates, with the position sent out over a serial interface; a 13-bit single-turn device divides one revolution into 8,192 distinct positions, and each additional bit doubles that; a multi-turn device adds a revolution counter, implemented either as a geared mechanical stage that is genuinely power-independent, or as an electronic counter kept alive by a battery or energy-harvesting element; the other consequence of counting is that errors persist, if electrical noise injects a spurious count, or a fast move outstrips the counter input, the controller does not find out, and the reported position simply drifts from reality
- Oman Seven Star, "Absolute vs Incremental Encoder: Selection Guide"
What we used
An incremental encoder produces a train of pulses as its shaft rotates, with two channels A and B offset by 90 electrical degrees (quadrature), and a third channel Z (index or marker) giving one pulse per revolution for homing; the limitation is fundamental: an incremental encoder only reports change, and when power is removed or a pulse is missed, the controller no longer knows the true position until the axis is homed again; an absolute encoder assigns a unique digital word to every shaft position, using a disc with several concentric code tracks (often Gray code, where only one bit changes between neighbouring positions); a 12-bit single-turn encoder resolves 4,096 positions per revolution, 13 bits gives 8,192, and 16 bits gives 65,536; absolute encoders dominate cranes, hoists, gates, stacker-reclaimers, rotary indexers, and any axis where an unexpected move after a power cut is unsafe or expensive
- Eltra Trade, "Incremental vs Absolute Encoders: Which One Should You Use?"
What we used
The main characteristic of every incremental encoder is data loss when power is turned off; an absolute encoder does not count pulses, it reports a unique digital code for each angular position of the shaft; if you turn off the machine, rotate the shaft by hand 90 degrees, and turn the power back on, the absolute encoder will immediately show the new correct position; an incremental encoder is very sensitive to electromagnetic interference, and running the encoder cable next to the VFD power cable without proper shielding will cause the controller to pick up false counts; an absolute encoder uses digital interfaces (SSI, EnDat, Profinet) that transmit data in packets, so if a packet is corrupted the controller generates an error but does not add an extra metre to the coordinate; incremental encoders have a PPR limit, while absolute encoders often offer much higher resolution, up to 20–24 bits
- Unseel, "Quadrature Encoder, A/B Channels, Direction and Position"
What we used
A quadrature encoder is an incremental position sensor with two output channels, A and B, whose square waves sit 90° apart; the phase order reveals the direction of motion, and counting the edges reveals position; decoding all four edges of each cycle multiplies the disc's line count by four, giving the cheap, high-resolution feedback inside every servo, robot joint, and scroll wheel; manufacturers rate an encoder in PPR (pulses per revolution, also called lines), and because the A/B pair has four distinguishable edges per cycle, a decoder that counts all of them resolves four counts for every line, CPR = PPR × decode factor, with decode factors of 1×, 2×, and 4×; a 1,000 PPR disc gives 4,000 CPR at 4× decoding, an angular resolution of 0.090°; a 5,000 PPR servo encoder gives 20,000 CPR; signaling is RS-422 differential, with A/Ā and B/B̄ pairs
- Design World, "FAQ: What do 1X, 2X, and 4X position encoding mean for incremental encoders?"
What we used
With X1 encoding, either the rising or the falling edge of channel A is counted; with X2, both the rising and falling edges of channel A are counted, doubling the resolution; with X4, both the rising and falling edges of both channels A and B are counted, quadrupling the number of pulses and increasing the resolution by four times
- Robotics Tomorrow, "What are Resolution, Accuracy & Repeatability"
What we used
Resolution defines the smallest measurable distance and therefore the smallest move increment; accuracy represents the deviation of the measured position from a calibrated standard; repeatability is a measure of the variance in successive position measurements at a specific location, typically specified as one least significant bit (LSB), which in terms of angle is the nominal size of one bin; precision is a system issue which can be dominated by mechanical errors, eccentricity, backlash, straightness, and flatness, and eccentricity and gearbox backlash are the most significant contributors to mechanical errors in rotary applications; as resolution increases and the bins get smaller, it becomes more likely for the true position to fall into the wrong bin if the accuracy stays the same; a servo dithers between two counts, and the higher the resolution, the smaller the dither
- ENC7, "Resolution, Accuracy and Repeatability: Three Specs That Are Easy to Confuse"
What we used
Resolution is how fine the scale is, accuracy is how far the reading is from the true value, repeatability is how consistent the readings are when the same position is reached again; accuracy is the maximum deviation, over one full revolution, between the reading and the true angle, and it comes from position errors of the lines on the disc, mounting eccentricity of the disc, bearing runout, and errors in the interpolation circuit; eccentricity is the most typical item, a disc eccentricity *e* at a reading radius *r* produces a once-per-revolution sinusoidal error of amplitude Δθ ≈ *e* ÷ *r*, and at a radius of 20 mm with an eccentricity of 3 µm, the error reaches about 31 arc-seconds at most, so on a 17-bit encoder, if the disc is mounted 3 µm off centre, the scale of the reading is 9.9 arc-seconds but the reading can be as much as 31 arc-seconds away from the true value, more than 3 counts; repeatability is often much better than accuracy, because systematic errors such as eccentricity and line-position error are the same every time and repeat unchanged when the same position is reached, and systematic errors can be measured and compensated
- Robo2U, "Rotary Encoders: Incremental, Absolute & Resolvers"
What we used
Resolution and accuracy are separate specs: a 14-bit magnetic encoder reports 16,384 counts per turn yet can be off by ±0.3°; optical leads on accuracy; the encoder is the only witness the controller can call, it is the sensor that tells the controller where the shaft is, and that is the whole job; resolution is not accuracy, and confusing the two is the single most common encoder mistake in robotics, a 14-bit magnetic encoder advertises 16,384 counts per turn (about 79 arc-seconds per count) but its real angular accuracy might be ±0.3° to ±0.5° (1,000 to 1,800 arc-seconds) once nonlinearity and mounting eccentricity are included, so the counts are precise and the angle is not; spec the encoder on the number that matches your control need, resolution for smooth velocity and low quantization noise, accuracy for absolute pointing and gear-train compensation, repeatability for return-to-home; the dirty secret is that velocity usually comes from differentiating position, and differentiation amplifies quantization noise brutally, a uniformly distributed quantization error has variance q²/12, and velocity noise scales as 1/(N·T), falling linearly with resolution and linearly with the differencing interval; the rule of thumb is at least 12-bit (4,096 CPR) resolution at the point you are controlling, and 17-bit+ (131,072 CPR) for direct-drive joints that must crawl smoothly; on a good drive, a state observer is doing more for low-speed smoothness than the encoder's headline bit count; the encoder also sets the commutation quality, a brushless motor commutated with a coarse or noisy angle wastes current as torque ripple and heat; the comparison table, optical ±1 to ±20 arc-sec accuracy, 28–32 bit max resolution, poor contamination tolerance, excellent stray-field immunity, high cost; magnetic (on-axis) ±0.1° to ±0.5°, 12–17 bit, excellent contamination, poor stray-field immunity, low cost; capacitive ±0.1° to ±0.2°, 12–14 bit, medium; inductive ±arc-minutes to ±0.05°, 18–22 bit, excellent on both, medium-high; resolver ±5 to ±20 arc-min, RDC-set 10–16 bit, excellent on both, medium; and the opinion, if you are building a robot arm and your knee-jerk is "magnetic, it's cheap and rugged", seriously look at inductive first, because you keep the ruggedness, lose the stray-field headache (your motor *is* a stray field), and gain a half-decimal-place of accuracy
- RobotForge, "Encoders: incremental, absolute, magnetic"
What we used
Incremental optical resolution is typically 100–10,000 PPR, and with 4× quadrature decoding you get 4× that in counts; strengths are cheap ($5–30), high-resolution, widely supported; absolute resolution is 12–22 bits per revolution (4,096 to 4 million positions), strengths are knowing position immediately on power-up, no missed counts, and multi-turn variants tracking total rotations, weaknesses are expensive ($50–500) and proprietary protocols adding complexity; magnetic resolution is 12–14 bits typical (4,096–16,384 positions), strengths are cheap ($3–20), no optics to fail, dust-immune, and contactless, weaknesses are lower resolution than optical and sensitivity to magnetic interference; the 2026 hobby default for BLDC motors is AS5048A or its successors, used in ODrive, MJBots Moteus, and SimpleFOC; magnetic encoders need the magnet within 0.5–1 mm of the chip and aligned to ~1°, and a slipped magnet after thermal cycling is a known failure; resolvers are absolute, military-grade, sine/cosine analog output, robust to vibration and temperature, common in aerospace and industrial servos; the decision matrix, hobby diff-drive wheels: incremental optical; BLDC motor commutation: magnetic; industrial arm joints: absolute (BiSS-C); aerospace, harsh environments: resolver
- Vanguard MagTech, "Motor Position Sensor Selection: Resolver vs Optical vs Magnetic Encoder"
What we used
Rotor-position feedback affects commutation angle, torque response, low-speed control, starting behaviour, efficiency, and fault detection; a device with excellent catalog accuracy can still underperform when shaft runout, target eccentricity, temperature drift, cable noise, and interface latency are added; selection should begin with the motor and drive operating envelope rather than the sensor technology name; the comparison, resolver: analog sine/cosine transformed by resolver-to-digital electronics, strong low-speed control when excitation and decoding are well designed, often selected for severe temperature environments, good sealed electromagnetic construction for contamination, differential analog signals requiring excitation and careful harness design, axial length and alignment, typical attraction is harsh-duty robustness and continuous absolute angle; optical: incremental or absolute digital position, strong with adequate resolution and interpolation, limited by optics, electronics, and materials on temperature, optical path sensitive if sealing is inadequate, digital outputs still requiring grounding and cable control, disk concentricity and air gap, typical attraction is high resolution and motion-control accuracy; magnetic: incremental or absolute digital/analog position, strong with suitable pole target, calibration, and bandwidth, limited by IC, magnet/target, and package on temperature, no optical path but target gap and debris still matter, must reject motor stray field and current-related disturbance, compact target and IC but strict magnetic geometry, typical attraction is compact absolute feedback and flexible integration
- Control Design, "High Noon for Encoder and Resolver"
What we used
Resolvers are inherently more rugged than encoders and less susceptible to electrical noise, and those are the only advantages; though resolvers theoretically have infinite resolution, in practice they are roughly equivalent to 1,000 PPR encoders; encoders have a glass disk that can shatter and contain electronics susceptible to heat and possibly vibration, but encoders are used all over the place; resolvers today are primarily relegated to a small corner of the motion control market, namely units built into the back of servo motors for feedback, and even there digital encoders are squeezing them out; all major drive and PLC manufacturers support encoder feedback, but only a handful support resolver feedback
- SpecForge, "Absolute Encoder: SSI, BiSS, EnDat, Multiturn"
What we used
An absolute encoder reports the actual angle of its shaft as a coded number, not as a stream of counts; every shaft position maps to a unique word, typically in Gray code, 17 bits for 131,072 positions per turn on an ordinary industrial unit, up to 23 bits, more than 8 million positions, on high-resolution optical designs; a multiturn variant additionally remembers the number of complete revolutions, by a geared counter, a battery, or a Wiegand-wire energy harvester, so an axis that traveled 37 turns and 90 degrees while powered down reports exactly that on wake-up; the buyer picks single-turn or multiturn by whether the machine can home after power-up, picks resolution and accuracy separately because they are not the same number, and then commits the controller side, the classic SSI serial link that any PLC input card reads, or one of the high-speed servo-class links, BiSS, EnDat, HIPERFACE, or an EtherCAT-native encoder profile, that push update rates into the tens of microseconds for closed servo loops; absolute disks are read in Gray code, in which adjacent positions differ in exactly one bit, so a sensor edge timing skew can delay a reading but cannot produce a wildly wrong value
- Parker Motion, "Absolute Encoders & Drive Compatibility"
What we used
An absolute encoder retains position information even when powered off; single-turn absolute encoders retain their orientation when turned off, and multi-turn absolute encoders retain their orientation AND the number of revolutions they have gone; BiSS-C absolute encoders are built into the Pseries motors and direct drive motors, are a lower-priced high-resolution absolute encoder option, and for multi-turn functionality require a battery inline (without it, single-turn absolute), with a multiturn resolution of 16 bit (65,535 turns); HIPERFACE DSL is available with a single-cable design carrying both power and feedback; the Heidenhain EQI1130 encoder is EnDat 2.1, 18 bit resolution (262,144 counts per revolution), multiturn with 4,096 revs
- Kübler, "Incremental Encoder Application Guide"
What we used
The resolution of incremental encoders is the number of pulses emitted within one revolution; incremental encoders typically have between 1 and 5,000 pulses, and high-resolution encoders up to 36,000 pulses per revolution; the pulse frequency must be considered from two points of view, the maximum allowed output frequency of the encoder (usually 300 kHz for standard types and up to 800 kHz for high-resolution encoders) and the input frequency of the pulse inputs of the controller or encoder card; the worked example: speed 3,000 rpm, resolution 1,000 ppr, pulse frequency = speed × resolution ÷ 60 = 50 kHz, and the pulse input of the controller must be able to process this frequency
- Schneider Electric, "Incremental Encoder Output"
What we used
The maximum frequency of the signal track is relevant for the evaluation unit of encoder signals, and the resolution of the encoder has to be selected in combination with the scaling factor in such a way that the output frequency of a signal track is not exceeded with the maximum velocity to be recorded; the maximum output frequency of a signal track is 1 MHz, and the guidance is to select the resolution so that an output frequency of almost 1 MHz is released at maximum machine velocity, with 800 kHz (20% reserve) at maximum velocity
- Torquety, "Servo Motor Velocity Control"
What we used
The velocity regulator requires a measured velocity feedback signal, and since most encoders measure position rather than velocity, the drive must derive velocity from the sequence of position measurements; the method used (and the trade-off between latency, noise, and low-speed resolution) significantly affects the velocity control performance; the worked example: a 1,000 PPR encoder (4,000 quadrature counts per revolution) running at 10 rpm produces 667 counts per second, and at a 4 kHz control loop rate that is 0.167 counts per loop cycle, because counts are integer values, the actual count per loop cycle alternates between 0 and 1, and the velocity computed from that alternates between 0 rpm and 60 rpm per count at the 4 kHz rate, which is extreme low-speed velocity noise; three methods, position differencing over a fixed time window (best at speeds where multiple counts occur per measurement interval), frequency measurement over a fixed count window (better low-speed velocity resolution, but variable latency proportional to 1/speed), and a combined hybrid method
- Thomas Helliez, "Encoder Velocity Estimation: Why Low-Speed Robots Jitter"
What we used
A robot creeps toward its docking station, the velocity loop runs at 1 kHz, but its encoder produces only about 130 counts per second, so most iterations report zero speed and an iteration that sees one count reports more than seven times the intended speed; before retuning the PID, check how velocity is estimated, count differences over a fixed window work well when that window contains enough counts, and timing the interval between encoder edges is often more useful at low speed; the worked example (2,048 decoded counts per wheel revolution, 0.020 m/s wheel-rim speed, 1 ms iteration): the nominal travel per count is 0.153 mm, the count rate is 130 counts per second, and the edge interval is 7.67 ms; the warning, if a datasheet specifies 512 electrical cycles per encoder-shaft revolution, x4 decoding produces 2,048 transitions per shaft revolution, but if the datasheet already specifies decoded counts, multiplying by four again introduces a fourfold scaling error, and the abbreviations PPR and CPR are insufficient without the manufacturer's definition
- Engineering.com, "Everything You Should Know About Functional Safety Encoders"
What we used
Functional safety encoders monitor a machine's operation and provide feedback if operation conditions reach a predefined unsafe state; the standards are IEC 61508 (the core functional safety standard, defining the Safety Integrity Level framework), IEC 61800-5-2 (encoders complying with it can be used for safe motor feedback, defining safety-rated drive functions), IEC 62061 (safety of machinery and functional safety based on EN 61508), and ISO 13849 (safety of machinery and safety-related parts of control systems, defining the Performance Level framework); encoders rated at SIL 3 (PLe) provide safe operation in set-up, production, and maintenance
- Design News, "Encoders Integrate Safety Technology"
What we used
Pepperl+Fuchs' RVS58S is the first and only ISO 13849-certified SinCos encoder specifically designed for use in safety-aligned systems up to SIL3 (IEC 61508, IEC 62061), performance level "e" (IEC 13849), and is suitable as a motor feedback system for safe drives in accordance with IEC 61800-5-2; the SinCos interface is compatible with existing safety PLC products, and since sine squared plus cosine squared equals one, the connected device can calculate whether the signal is genuine, uninterrupted, and reliable, and can detect a wire break, a key issue and limitation with incremental encoders; the encoders offer 1,024 or 2,048 signal periods, deliver a 1 Vpp output signal up to a maximum output frequency of 200 kHz, and are rated to 115 °C
- Renishaw, "RESOLUTE FS / TONiC FS Functional Safety Encoders"
What we used
The RESOLUTE Functional Safety (RESOLUTE FS) optical absolute encoder and the TONiC Functional Safety (TONiC FS) incremental open optical encoder system are certified to ISO 13849 Category 3 PLd, IEC 61508 SIL2, and IEC 61800-5-2 SIL2; full functional safety certification is required to guarantee the safe use of functions that include safe stops 1 and 2, safe operating stop (SOS), and safe limited speed (SLS), all of which require functionally safe encoder feedback; the TONiC FS provides safe position data supporting SS1, SS2, SOS, Safe Limited Acceleration, Safe Acceleration Range, Safe Limited Speed, Safe Speed Range, Safely Limited Position, Safely Limited Increment, Safe Direction, and Safe Speed Monitor; the system must be installed and operated in accordance with the instructions in the installation guide, or SIL2 and/or PLd may not be achieved and the certification is invalidated
- Typhoon HIL, "Providing absolute machine encoder position over a standardized protocol"
What we used
Absolute encoders provide a unique machine rotor position and a number of machine rotor revolutions over SSI, BiSS, or EnDat protocols; the HIL device accepts clock frequencies up to the protocol specification maximum (SSI: 2 MHz, BiSS-C: 10 MHz)
- Article #20, "Choosing the Motor: Torque Decides, Not RPM" (Senova, 2026-11-19)
What we used
The motor the encoder closes the loop on; the bench arm's joints (shoulder 126.8 N·m with margin, elbow 40.7, wrist 11.0), the 100:1 planetary at 93% efficiency, the motor at 1.36 N·m and 1,500 RPM, the joint at 15 RPM; the smart servo positions to ±0.088°; the servo is a motor plus feedback plus a correcting drive, sold as one package; the FOC's two current loops run at 10–40 kHz; position accuracy is encoder-limited, better than 0.01° with a good encoder
- Article #17, "Sensor Calibration and Drift" (Senova, 2026-11-19)
What we used
The encoder is the second sensor that checks the gyro; a gyroscope measures rate, not angle, and the encoder is the absolute reference the fusion filter anchors to
- Article #3, "PID Tuning, Simulation First" (Senova, 2026-11-19)
What we used
The PID loop whose noise floor the encoder's resolution sets
- Article #21, "The Gearbox Is the Joint" (Senova, 2026-11-19)
What we used
The harmonic wind-up of 10.8 mm at the 0.85 m reach under the 126.8 N·m shoulder load, 43 times its backlash
- Article #22, "The Arm That Feels" (Senova, 2026-11-19)
What we used
The peak contact force of 1,159 N under position control, and the safety limits the encoder's trusted position must respect