Rotary encoder

A rotary encoder is an electromechanical sensor that measures the angular position, direction, and speed of a rotating shaft and converts that motion into an electrical signal a controller can read.^[1] It is the part that lets a motor drive, a CNC machine, or a robot joint know exactly where a shaft has turned to, rather than only how much current a motor is drawing. This article covers the physical position and motion sensor used in industrial automation and robotics, not the encoder, or the encoder half of an encoder-decoder architecture, used in transformer and sequence-to-sequence neural networks; AI Wiki covers that unrelated software concept in a separate article on machine learning encoders.

In brief: a rotary encoder is a joint's sense of where it is. A motor alone only knows its current draw; the encoder on its shaft, or on an actuator's output after a gearbox, tells the control loop the real angle, so the system can hold a pose, walk, or grasp an object without guessing.

How it works

A rotary encoder pairs a moving element on the shaft (a patterned disc, a small magnet, or a conductive target) with a stationary read head that senses the pattern as it passes.^[1] As the shaft turns, the read head detects periodic changes (light through slots, a rotating magnetic field, a shifting capacitance, or a modulated electromagnetic coupling) and converts them into an electrical output, in one of two forms. An incremental encoder emits a pulse train, commonly two square waves 90 degrees out of phase called quadrature signals, that a downstream counter tallies to track movement since it was last reset.^[2][3] An absolute encoder instead outputs a unique digital code for every distinct angular position, so a controller reads the shaft's exact angle instantly at power-up, with no counting or prior reference needed.^[2][3] Both output styles can be built on any of the sensing technologies below; how an encoder senses motion and what it remembers about position are independent design choices.

Types and variants

Encoders are usually described along three independent axes: how the mechanism moves, what physical effect the encoder uses to sense that movement, and what the encoder remembers about position when power is removed.

Motion: rotary versus linear

Most encoders used in robotics are rotary, sensing angle around an axis. This fits most robot joints and motor shafts, which rotate rather than slide: knees, elbows, wrists, and the motors driving them all suit a device that reads angle, and rotary encoders show up on motor shafts, inside gearboxes, and at joint outputs.^[1] The related category is the linear encoder, which measures straight-line displacement along a scale. Linear encoders dominate high-precision machine tools and appear in robots built around linear actuators such as screw drives, but they are far less common across the rotating joints that make up most of a humanoid robot's degrees of freedom.^[1][4] Many linear-motion robotics systems instead infer linear position from a rotary encoder on the driving screw, which is usually smaller, cheaper, and easier to seal than a dedicated linear scale.

Sensing technology

The physical effect an encoder uses to detect position is largely independent of whether it is incremental or absolute. Four sensing technologies are in common use, each trading resolution, robustness, size, and cost differently.^[4][5][6]

Optical encoders shine light, typically from an LED, through slots in a glass or plastic disc, or reflect it off a patterned track, onto a photodetector. Because the pattern can be etched at very fine pitch, optical encoders deliver the highest resolution and accuracy of the four technologies, in some designs resolving tens of thousands of counts per revolution.^[4][5] The tradeoff is fragility: dust, oil film, condensation, and vibration can all disrupt the light path, so optical encoders generally need a sealed housing or a clean environment, which is why they mostly appear inside enclosed motor modules rather than at an exposed outer joint.^[4]

Magnetic encoders replace the optical disc with a small permanent magnet, often a diametrically magnetized disc or a multi-pole ring, and read its field with a Hall-effect or magnetoresistive chip mounted nearby.^[4][5] With no glass disc to crack and no light path to block, magnetic encoders tolerate dirt, oil, moisture, and shock far better than optical designs, at the cost of somewhat lower resolution and accuracy; stray fields from nearby motors, ferrous debris, and temperature swings can still degrade a reading.^[5] Magnetic sensing has become the preferred choice specifically for humanoid and legged-robot joints, a point discussed further below.

Capacitive encoders use two closely spaced sets of electrodes, typically etched directly onto printed circuit boards (one rotor, one stator), and measure the changing capacitance between them as the rotor turns.^[6][7] Because the sensing pattern is just PCB copper rather than a magnet or an optical grating, capacitive encoders can be built thin and achieve high resolution in a compact package, and they are inherently wear-free since the plates never touch.^[6][7] They remain a smaller niche, more sensitive to humidity and electromagnetic interference than the alternatives, but suppliers including CUI Devices and Netzer Precision market capacitive designs for tight, weight-constrained spaces such as robotic hand and finger joints, where a dexterous hand has little room for a full-size sensor.^[6][7]

Inductive encoders drive an alternating current through a coil to generate an electromagnetic field, then read how a passive rotor (a shaped piece of metal or a printed conductive target) disturbs that field through induced eddy currents as it moves.^[8][9] HEIDENHAIN, one of the few manufacturers with deep expertise here, positions inductive sensing between optical and magnetic on most measures: above-average accuracy alongside strong tolerance for vibration and contamination, without an optical disc's fragility.^[4] TE Connectivity sells a related eddy-current position sensor descended from the resolver, an older analog electromagnetic position sensor long used in aerospace and motor feedback.^[9] Novanta's Celera Motion division acquired the inductive-encoder specialist Zettlex in 2018 and now markets PCB-based inductive encoders for industrial and medical robotics, a non-contact option that keeps working where magnetic and capacitive designs struggle.^[8] Because few companies have the process expertise to build them, inductive encoders remain a smaller share of the market than optical or magnetic designs despite generally favorable performance.^[4]

Scroll sideways for more →

Memory: incremental versus absolute, single-turn versus multi-turn

Separately from how it senses motion, an encoder differs in what it remembers about position once power is removed.

Incremental encoders only report change: they count pulses while powered and have no idea of the shaft's actual angle, so if power is cut, that count is lost. On restart, a system built around one typically must run a homing routine, driving the shaft to a limit switch or index mark to re-establish a zero reference, before position can be trusted again.^[2][3] Incremental encoders are simple, fast, and inexpensive, and remain common for measuring raw motor-shaft velocity, where a controller cares about speed rather than absolute angle.^[2]

Absolute encoders assign a unique code to every distinct angular position, so they report the shaft's exact position the instant they are powered, with no motion and no counting required.^[2][3] Within a single revolution this is a single-turn absolute encoder; because that code necessarily repeats every 360 degrees, tracking position across many revolutions, needed for a highly geared or screw-driven axis, requires a multi-turn absolute encoder that adds a way to count whole turns on top of the within-turn angle.^[10][11]

Multi-turn tracking has historically been done three ways.^[10][12] The oldest gears a second, slower code disc to the primary disc so it advances one position per full revolution, a purely mechanical solution that adds moving parts, size, and wear points. A more compact alternative keeps an electronic turn counter alive with a small battery, incrementing while the system is unpowered, at the cost of periodic replacement. The newest, increasingly common approach exploits the Wiegand effect or similar energy harvesting: shaft rotation itself generates a brief but strong pulse through a specially treated wire, powering a turn counter that writes to non-volatile memory with no battery and no gear train.^[12] Kübler markets multi-turn magnetic encoders built this way, describing them as gearless and batteryless.^[13]

Scroll sideways for more →

Key evaluation criteria: resolution, accuracy, and repeatability

Encoders are compared on three related but distinct properties, and conflating them is a common mistake.^[14][15][16]

Resolution is the smallest change in angle the encoder can distinguish, expressed as pulses or counts per revolution for incremental designs, or as a bit depth for absolute designs (a 16-bit encoder resolves one turn into 65,536 steps). Higher resolution matters most for smooth control at low speed, where coarse steps would otherwise show up as jerky motion.^[14][16]

Accuracy is how close the reported angle is to the shaft's true physical angle. A magnetic encoder can carry very high resolution, millions of interpolated counts, while still being less accurate than a coarser optical encoder, because small distortions in the magnet's field or slight misalignment of the read head introduce error that resolution alone does not fix.^[15][16] Resolution and accuracy are independent: two encoders can share a resolution figure yet differ substantially in accuracy.

Repeatability, sometimes called precision, is the encoder's ability to report the same value every time the shaft returns to the same physical position, and is typically several times smaller and more consistent than absolute accuracy.^[15][16] In robotics it often matters more than raw accuracy, since a small, consistent offset can be measured once and corrected in software, while random error cannot.

These properties are not needed uniformly across a humanoid robot's joints: a finger or hand joint used for fine manipulation typically needs much higher resolution than a hip or knee joint, where torque, durability, and speed dominate instead.^[17][21] Guidance for demanding manipulation joints suggests resolutions of roughly 16-bit to 23-bit or higher (65,536 to more than 8 million counts per revolution), while a leg joint can often accept coarser steps for a smaller, tougher, cheaper sensor.^[21]

Use in humanoid and legged robots

Humanoid and legged robots place demands on rotary encoders that most industrial motion-control applications do not share.

Why absolute encoders are effectively mandatory

A wheeled or fixed industrial robot can often afford to lose track of position when it loses power, since it can safely run a homing move before resuming work. A legged or humanoid robot generally cannot: it may be powered off, or lose power unexpectedly, in almost any pose, crouched, mid-stride, or with an arm extended, and a bump or gravity can shift an unpowered joint before power returns.^[17][18] If the controller does not know the true angle of every leg and torso joint the instant it wakes up, it cannot safely work out where the robot's center of mass and feet are relative to each other, information needed to stand up, balance, or move a single joint without a collision or a fall.^[17][19] Driving joints to a hard stop to "find home," the normal recovery for an incremental encoder, is exactly the kind of blind, uncontrolled motion that is dangerous on a machine with a humanoid's mass and reach. For this reason, virtually every joint on a modern humanoid or legged robot uses an absolute rather than an incremental encoder, despite the higher cost.^[3][17]

A documented example is PAL Robotics' REEM-C, a human-sized biped with up to 40 degrees of freedom, whose knee, wrist, and elbow joints use Renishaw's AksIM and Orbis absolute magnetic encoders. Direct joint-angle feedback lets the control system keep the robot's zero moment point, the standard balance criterion for bipedal walking, within the support area of its feet, rather than relying on an assumed or homed position.^[18]

Why magnetic absolute encoders are becoming the default

Among the four sensing technologies, magnetic designs have become the preferred choice specifically for humanoid and legged-robot joints, for a combination of reasons rather than any single one.^[4][5][17] Magnetic encoders tolerate the dust, metal shavings, lubricant, and impact loads around gearboxes far better than an optical disc, while needing less sealing and fewer precision alignment steps, which lowers cost.^[5][17] They can also be built as thin, hollow-shaft, off-axis rings rather than an on-axis shaft-end mount, letting designers route cables or a joint's output bearing through the encoder's own center, which matters when packing dozens of actuators into a human-sized frame.^[20] The same magnetic chips used for position sensing can typically also supply the commutation signal a brushless DC motor needs, so one absolute magnetic encoder can replace the separate Hall-effect sensors a servo motor would otherwise carry.^[26] They still trail optical and inductive designs on raw resolution and accuracy, which is why some designers reserve those technologies for joints where the extra precision is worth the added cost and fragility.^[4]

The dual-encoder practice and the sim-to-real gap

A motor's own shaft encoder only reports what the motor is doing, not what the joint it drives is actually doing. Between the motor and the load sits a gearbox, commonly a harmonic drive or cycloidal drive in humanoid joints, and any real gearbox has some backlash, torsional compliance, or elastic windup under load, so a controller trusting only the motor-side reading is effectively guessing at the true joint angle once the gearbox is loaded.^[17][22][23]

The common fix is a second encoder at the joint's output, after the gearbox. With both readings available, the controller can directly measure and correct for backlash and transmission compliance instead of assuming it away, improving position accuracy and cutting output speed fluctuation compared with a motor-side encoder alone; if the transmission itself slips or fails, the output-side encoder still reports the load's true position.^[22][23] ZeroErr builds its eRob actuator modules with absolute encoders at both the motor and the harmonic-drive output, publishing repeatability of about 7 arcseconds and absolute accuracy of about 15 arcseconds, explicitly to offset backlash and wear in the reducer.^[24]

This connects to sim-to-real transfer, the gap between a control policy's behavior in simulation and on physical hardware. Simulators typically model a joint as a rigid link between commanded and actual angle, while real actuators add friction, backlash, and compliance that distort that relationship.^[25] Research on tendon-driven robot hands has found that reading joint angle directly at the output, rather than inferring it through a stretchy, frictional cable transmission, gives a truer proprioceptive signal and narrows that gap; the same logic applies to a geared humanoid limb, where a joint-side encoder gives a reinforcement-learning pipeline ground truth instead of a motor-side estimate corrupted by whatever sits in between.^[25]

Some designs skip the second encoder: a low-ratio quasi-direct drive, popularized on quadruped legs and used on some humanoid limbs, has little backlash to correct for, so a single motor-side encoder is often adequate. Higher-reduction harmonic and cycloidal joints are typically where the dual-encoder architecture appears instead.^[22][23]

Encoders alongside other proprioceptive sensors

A rotary encoder reports a single joint's own rotation; it does not measure the robot's overall orientation or acceleration in space, which instead falls to an inertial measurement unit on the torso or head. Joint encoders build a model of the robot's pose from the inside out, while an IMU measures how the whole body tips or accelerates in the world, and control software typically fuses both for balance and locomotion.

Some designs also estimate joint torque from motor current and encoder data rather than adding a dedicated force-torque sensor at every joint. Current-based estimates are cheaper, but research comparing the two has found dedicated torque sensors measurably more accurate, since friction and elasticity in the gearbox are hard to model from current alone.^[27] That gap is one reason some designs still reserve a dedicated torque or tactile sensing sensor for a wrist or end effector, where contact force matters most.

Suppliers and landscape

Encoder manufacturing splits roughly into complete assemblies (a housed sensor plus electronics) and bare encoder chips that a motor or robot maker integrates onto its own board. Both categories appear among the companies most often named in robotics and industrial-automation encoder catalogs.

Heidenhain and Renishaw anchor the high end of the industry. Heidenhain, refounded in Traunreut, Germany, in 1948 out of an 1889 Berlin scale-making business, began building photoelectric optical position encoders in the early 1950s and is still known for machine-tool-grade optical and angle encoders.^[28] Renishaw built its name on metrology touch-trigger probes before adding encoder readheads in 1989; its AksIM and Orbis magnetic encoders are the ones documented in PAL Robotics' REEM-C humanoid.^[18][29]

Broadcom sells the AEAT family of magnetic encoder chips (10-bit to 18-bit programmable angular resolution, on-axis and off-axis), inherited through Avago Technologies' 2016 acquisition of the earlier Broadcom Corporation, after which Avago took the Broadcom name.^[30][31] Baumer began as a Swiss microswitch maker in 1952 and now sells optical, magnetic, inductive, and capacitive encoders, including aerospace-qualified designs for Airbus.^[32][33] SICK, founded in 1946 for photoelectric safety sensors, added rotary encoders through its acquisition of Stegmann; the company reported revenue of roughly 2.1 billion euros and about 11,800 employees in 2024.^[34][35]

Dynapar's brand portfolio includes Hengstler, NorthStar, and Harowe; owned by Fortive since 2016, it joined the newly spun-off Ralliant Corporation (NYSE: RAL) in June 2025.^[36][37] Kübler (Fritz Kübler GmbH), a family-owned German company that entered encoders in 1989, sells Sendix magnetic multi-turn encoders using gearless, batteryless energy-harvesting technology instead of a battery or gear train.^[13][38]

ZeroErr, based in Shenzhen, China, is a newer entrant built for robotics: its eRob modules combine a motor, harmonic reducer, and dual absolute encoders in one housing, marketed at collaborative, surgical, and humanoid or bionic robot builders.^[24][39] Melexis, a Belgian semiconductor company, supplies Hall-effect position-sensor chips such as the MLX90363, meant to be integrated onto a customer's own motor or joint board.^[40] iC-Haus, a German chip maker, supplies optical and Hall-based encoder chips to other equipment makers, including designs using blue LEDs for better resolution.^[41] CUI Devices, rebranded as Same Sky in 2024, sells the AMT modular encoder family, built on capacitive rather than magnetic or optical sensing and marketed as immune to dirt, dust, and oil.^[42][43]

Scroll sideways for more →

See also

• Actuator
• Servo motor
• Brushless DC motor
• Harmonic drive
• Inertial measurement unit
• Sim-to-real transfer
• Degrees of freedom
• Humanoid robots

References

1. Rotary encoder, Wikipedia ↩
2. Rotary Encoder Options: Absolute or Incremental?, DigiKey ↩
3. What is the difference between an incremental and an absolute encoder?, US Digital ↩
4. Inductive vs. optical vs. magnetic encoders: how to choose, HEIDENHAIN ↩
5. Magnetic vs Optical Encoder, POSITAL ↩
6. Capacitive, Magnetic, and Optical Encoders: Comparing the Technologies, Same Sky (CUI Devices) ↩
7. High-Resolution Joint Encoders for Industrial Robotics, Netzer Precision ↩
8. Zettlex Inductive Position Sensors, Celera Motion (Novanta) ↩
9. Eddy Current Resolver / Motor Position Sensor, TE Connectivity ↩
10. Single-Turn Encoders vs Multi-Turn Encoders, Dynapar ↩
11. The difference between single-turn and multi-turn encoders, HEIDENHAIN ↩
12. FAQ: What are traditional and newer multi-turn measuring systems for encoders?, Motion Control Tips ↩
13. Sendix M3681 encoder, absolute multiturn, Kübler Group ↩
14. Accuracy, Repeatability and Resolution: What's the Difference?, HEIDENHAIN ↩
15. Encoder resolution, accuracy and repeatability: What's the difference?, RLS ↩
16. Encoder Resolution, Encoder Accuracy & Repeatability, Dynapar ↩
17. Humanoid robots, RLS ↩
18. PAL Robotics integrates magnetic encoder technology into robots to achieve balance, Renishaw ↩
19. Absolute Encoders in Automation, US Digital ↩
20. The Application of Off-Axis Absolute Encoders and Magnetic Rings in Robot Joint Modules, OTV Precision Sensing ↩
21. Choosing Encoders for Humanoid Robot Joints, FLUX GmbH ↩
22. Vibration Suppression of Collaborative Robot Servo with Dual Encoders, Applied Sciences (MDPI) ↩
23. A Complete Guide to Choosing Single-Encoder vs. Dual-Encoder Robot Joint Actuators, Honpine ↩
24. eRob 90I rotary actuator, ZeroErr ↩
25. Actuator Reality Shaping for Zero-Shot Sim-to-Real Robot Learning, arXiv ↩
26. Feedback Devices: Exploring Hall-Effect Sensors and Resolvers, Kollmorgen ↩
27. Physics-Informed Neural Networks with Unscented Kalman Filter for Sensorless Joint Torque Estimation in Humanoid Robots, arXiv ↩
28. Milestones in the history of HEIDENHAIN, HEIDENHAIN ↩
29. Renishaw: Our history, Renishaw ↩
30. Avago Technologies to Acquire Broadcom for 37 Billion Dollars, Broadcom investor relations ↩
31. AEAT-9922: Programmable Angular Magnetic Encoder IC, Broadcom ↩
32. History, Baumer ↩
33. Magnetic Encoders from Baumer Highlighted at SPS IPC Drives, Magnetics Magazine ↩
34. SICK's company history, SICK ↩
35. Sick AG, Wikipedia ↩
36. About Us, Dynapar ↩
37. Fortive Finalizes Ralliant Spin-Off, Nasdaq / Zacks ↩
38. Kübler Group, worldwide, Kübler Group ↩
39. About us, ZeroErr ↩
40. MLX90363: Triaxis mainstream rotary, linear and joystick position sensor IC, Melexis ↩
41. World of Mini Encoder ICs, iC-Haus ↩
42. AMT Modular Encoders, Same Sky (CUI Devices) ↩
43. CUI Devices has a New Name, Say Hello to Same Sky, Same Sky ↩