Design Book › Sensors

Sensors

For each sensor, a designer must decide what state it proves, how it fails, and what bus and rate carry it. Do these first.

  1. List the states you must know directly. Critical joint states should be measured directly, for example with an output encoder [1].
  2. Choose the family from the failure you can tolerate: dust, stray fields, sunlight, drift.
  3. Check the conditions on any headline number. Range and accuracy are quoted at a stated reflectivity or temperature.
  4. Sample at equal intervals. Velocity is derived from position, so small timing errors become velocity errors [2].

After you choose a part, timing, calibration, mounting and limits are in Sensor integration.

The sensor guidelines cover integration, not selection; see Sensor integration. So this chapter rests on part data. Prices below are public quantity 1 rows from the price table, and EUR or GBP rows use its USD conversion. Where the only values for a part come from a source that does not let us copy them, the text names the part without the numbers.

Encoders

Magnet mounting and load accuracy guidelines are in Sensor integration.

Encoders
FamilyHow it worksGood atWeak atParts
MagneticA magnet read by Hall or magnetoresistive sensingCheap; tolerates dust, oil and vibration; absolute output removes homingSet by magnet quality and alignment; stray fields degrade accuracyMT6701, AEAT-8800, Orbis BRD16, AS5600, AS5047P
CapacitiveA rotor modulates field coupling between PCB electrodesImmune to magnetic fields; thin, hollow shaftFew vendors; pricing on requestNetzer VLX
OpticalLight reads a patterned discHighest resolution and accuracy of the common familiesContamination, shock and condensation; usually bigger and costlierAEDM-5810, AEDT-9810
Hall quadratureHall sensors on a magnetic disc, for gearmotorsSimple kitLow countPololu 20 CPR kit
Encoders
PartTypeValues on filePrice, USD
AEDT-9810-Z00Optical, incremental5000 pulses per turn per channel; 26 mm code wheel; 5 g (claims, ODRI actuator module)none
AS5600 breakoutMagnetic angle sensor3 mm sense distance (claim)5.95

The database also holds resolution, accuracy and speed rows for the MT6701, AEAT-8800-Q24, Orbis BRD16 and Pololu kit. Their sources do not let us copy them, so read the maker's datasheet. Bare magnetic ICs have quantity 1 price rows, but those are not shown for the same reason (encoder category). At the high end, a capacitive encoder reaches finer accuracy and resolution than a magnetic IC.

Encoder accuracy does not remove joint error. The STS3215 encoder step is a small fraction of a degree, yet a bench test measured about 1 deg of backlash behind it [3]. Sensor integration compares the step with the backlash. On a strain wave joint, count hysteresis and wind-up as well [4], [5].

IMUs

MEMS parts are small and cheap. Their limits are bias drift, magnetometer disturbance from motors and steel, and vibration. A fusion part returns orientation without host-side filtering, but gyro bias still drifts without a reference. See the IMU category. Mounting, vibration and noise budgets are in Sensor integration.

IMUs
PartTypeRangeOtherPrice, USD
ICM-20649 breakout6-axis, wide range30 g, 4000 dps14.95
ISM330DHCX board6-axis16 g, 4000 dps, selectable maximum1 x 1 in boardnone
LSM6DS3TR-C breakout6-axis2000 dps1.6 kHz output rate, 1.7 g9.95
ICM-20948 breakout9-axis2000 dps; compass to 4900 uT19.95
BNO085 breakout9-axis fusionnot listed2.5 g29.5
BNO086 breakout9-axis fusionnot listedRotation vector error 3.5 deg dynamic, 2.0 deg static; game rotation vector drift 0.5 deg/min (SparkFun, from the BNO086 datasheet)not shown

Compare IMUs by noise density; Sensor integration shows why. The database holds noise density rows for bare chips such as the ICM-42688-P, BMI270, BMI323 and BMI088, and range rows for the BNO085 chip. Their sources do not let us copy them, so read the datasheet. The BMI088 is meant for high-vibration mounts.

IMU boards with a public quantity 1 price: 6-axis $5.69 / $12.50 / $19.95 over seven boards, 9-axis $19.95 and $22.50, fusion $29.50 and $34.95. Bare IMU chips have price rows too, but their sources do not let us copy them, so those prices are not shown.

Force and torque

A 6-axis sensor gives direct contact force. It needs temperature and gravity compensation. Overload can damage the flexure. Most 6-axis sensors have no price row: vendors quote on request. Two DOBOT KWR75B listings have one (transducer, cobot sensor), but the seller does not let us copy the price. See the force-torque category. Overload, bias and temperature guidelines are in Sensor integration.

The one 6-axis sensor with public values is the Robotiq FT 300-S, a strain gauge sensor: 300 N and 30 Nm range, 100 Hz data rate, 0.44 kg and 500 % overload. The database also covers an optical sensor (OnRobot HEX), a capacitive-style one (Bota Rokubi) and ATI strain gauge models (Mini40, Delta SI-165-15). Their values are not shown, because their sources do not let us copy them.

Silicon strain gauges, as on the Nano17, give fine resolution. They cost more, and need a calibration and a DAQ or controller. The ATI strain gauge range runs from fingertip sizes to large wrist sensors. Fz is often rated differently from Fx, so read each axis in the maker's table. Joint torque sensing is a separate route, and the cobot class lists it for contact detection.

Tactile

Tactile
FamilyHow it worksExamplesLimits
Force-sensing resistorPolymer film; resistance falls with forceSquare FSR to 49 N, 0.55 mm thick, 38 mm wide (claim). Small FSR to 10 N on a 4 mm pad (claim). Round FSR 0.3 to 10 N, $3.95. Tekscan FlexiForce A201 in 1 lb, 25 lb and 100 lb models: values not shownRepeatability and drift limit accuracy. Single point.
Magnetic skinMagnetised elastomer over magnetometersAnySkin: its authors report a smaller loss than ReSkin when a trained model moves to a swapped skin, on one plug insertion taskSensitive to external magnets; lower spatial resolution
Vision-basedCamera behind a soft gelDIGIT, an open design whose values are not shown, and GelSight Mini, a commercial one with no value on fileGel wears; needs camera bandwidth

Cameras

Choose by the depth method first. All value rows link to the part. See the camera category. Shutter, minimum depth and hand-eye guidelines are in Sensor integration.

Cameras
MethodGood atWeak atParts
Active or passive stereoOutdoors, wide choice of baselinesDepth error grows with range; baseline sets minimum rangeD435, ZED 2i, OAK-D Pro, Gemini 335
Structured lightDense depth on plain indoor surfacesSunlight washes it out; cameras can interfereAstra 2
Time of flightDepth independent of colour and textureMultipath and sunlightFemto Bolt
Cameras
PartBaselineDepth rangeDepth shutterOther
D435not shown0.2 to 10 mglobal90 deg diagonal field of view
D415not shown0.5 to 3 m ideal; about 45 cm minimum at full resolutionrolling99 x 20 x 23 mm
D455not shown0.6 to 6 m ideal; about 52 cm minimum at full resolutionglobal124 x 26 x 29 mm
OAK-D75 mmnot shownglobal6 to 7.5 W typical, 2.5 W idle
ZED 2i120 mm0.3 to 20 mnot listed166 g; 175 x 30 x 33 mm

The RealSense and ZED rows come from published papers that tabulate the makers' datasheets. The OAK-D row comes from the Luxonis hardware repository. The database also holds values for the D405, Gemini 335, ZED Mini, ZED X Nano, ZED X Mini, ZED X, Astra 2 and Femto Bolt, and for the RealSense baselines. Their sources do not let us copy them. Depth camera prices are not shown for the same reason.

For colour without depth: Camera Module 3 is 12 MP, and Adafruit sells it at $29.25. The Global Shutter Camera is 1.6 MP with 3.45 um pixels. Machine vision cameras run 0.4 / 5 / 31 MP and 5 / 48.5 / 522 FPS at $390 / $1419 / $10633.50. The class note says from about 540 USD, and the price table holds lower rows. Trust the rows.

Give each camera an optical frame in the model [6]. Verify camera assignments after any replug, since indices can change [7].

LiDAR

Do not compare range numbers without the reflectivity. The RPLIDAR S2E profile gives 30 m, but 10 m at 10 percent reflectivity. Unitree L2 lists 30 m at 90 percent. These are different tests. See the lidar category. Blind zone and range condition guidelines are in Sensor integration.

LiDAR
PartTypeRangeOther
LD-062D, 360 deg12 m4500 samples per second; IPX4
RPLIDAR S2E2D ToF30 m; 10 m at 10 percent32000 points per second; 0.05 m minimum
SF45/B2D, up to 350 deg50 mabout 50 g
picoScan1502D45 m (claim)0.05 m minimum, 20 Hz (claim)
Hesai XT323D spinning120 m0.05 m minimum; 0.18 deg step; 10 Hz
VLS-1283D spinning, 128 channels200 m (claim)1 m minimum (claim)
multiScan1363D, -22.5 to +42.5 deg vertical60 m (claim)0.05 m minimum, 20 Hz (claim)
ML-30s+Solid-state45 m (claim)0.2 m minimum (claim)
Unitree L2Non-repetitive30 m at 90 percent (claim)64000 points per second, 230 g, 0.05 m blind zone (claims); $419

The LD-06 and SF45/B rows come from the ArduPilot wiki. The S2E, XT32, VLS-128, SICK and ZVISION rows come from the sensor profiles that ship with Isaac Sim, which cite the makers' datasheets. The L2 row is the maker's claim.

A 2D unit sees one plane only. Triangulation units such as RPLIDAR A2M8 have a short range, and sunlight can disturb them. Spinning units such as RPLIDAR A1 have moving parts that wear. Non-repetitive units such as Mid-360 need algorithms built for the pattern. The database also holds values for the RPLIDAR A3, LD19, Hokuyo UST-10LX, Hokuyo UTM-30LX, Hesai JT16, Hesai QT128, Ouster OS0, OS1, OS2, Livox Avia and Mid-70. Their sources do not let us copy them. Most lidar price rows are EUR reseller listings with the same limit, so the L2 is the only lidar with a public price.

Range sensors

Ambient light and ToF calibration guidelines are in Sensor integration.

Range sensors
PartTypeRangeOtherPrice, USD
HC-SR04Ultrasonic2 to 400 cm3 mm accuracy, best casenone
VL6180X breakoutLaser ToF100 mm (claim)13.95
VL53L4CD breakoutLaser ToF1300 mm (claim)14.95
VL53L1X breakoutLaser ToF4000 mm (claim)50 Hz maximum update (claim)14.95
VL53L5CX mini imagerMultizone ToF4000 mm64 zones, 63 deg diagonal, 15 Hznot shown
TFmini-SSingle-point lidar0.1 to 12 m at 90 percent; 7 m at 10 percent1000 Hz, 2 deg, 5 gnone
TFmini PlusSingle-point lidar0.1 to 12 m at 90 percent; 4 m at 10 percent1000 Hz, 3.6 deg, 12 gnone

The TFmini rows come from the PX4 documentation. The database also holds values for the GP2Y0A21, the Pololu VL53L1X, VL53L5CX and VL53L7CX carriers, and the ranges of the TF-Luna and TF-NOVA. Their sources do not let us copy them.

Sources differ for the same chip. Adafruit lists 1200 mm for its VL53L0X board, and the ArduPilot wiki gives 2 m for the Pololu carrier. Range varies with ambient light and reflectivity, so test on your target.

Current sensing

A shunt monitor digitises the voltage across a resistor. A Hall IC measures the field around a conductor and isolates the load.

Current sensing
PartTypeKey valuesNote
INA226I2C shunt monitor36 V busShunt dissipates power and is not isolated
INA228I2C shunt monitor85 V bus
INA219I2C shunt monitor26 V bus
INA180A1Analog amplifier20 V/V gain
ACS712Isolated Hall IC5 V supply
ACS724LLCTR-10ABHall IC10 A, 200 mV/A, 5 V supply

Accuracy, resolution, bandwidth and isolation rows for the INA219, INA226, INA228, INA180 and ACS712 are on file, but their sources do not let us copy them. Read the datasheet before you trust a current reading.

At quantity 1, the three I2C power monitor boards with a public price cost $9.95 / $9.95 / $14.95 (current sensor category). Bare current sense amplifier prices are not shown. Motor current is also a torque estimate, but only at low ratio [8].

Sources

Each guideline in this chapter with its source, the date it was read and the exact words of the source. Level: must means breaking it causes failure or harm, should is good practice, guide is a rule of thumb.

  1. Provide direct indication of critical mechanism states. Level: should. standards.nasa.gov 2026-11-08“[DDMR 26] A direct indication of the critical states of each mechanism shall be provided.”
  2. Take position measurements at exact equal intervals, since velocity is derived from them and small timing errors become velocity errors. Level: should. ethercat.org 2026-11-08
  3. A bench test of the STS3215 measured about 1° of encoder backlash against a ≤0.5° datasheet value. Level: guide. github.com 2026-11-08
  4. Model strain wave torsional stiffness as three segments K1, K2, K3 across T1 and T2, not as one constant. 3 Level: should. harmonicdrive.net 2026-11-08
  5. Typical hysteresis loss is 3 arc min at ratio 30; use table 44 for the applicable size and ratio. 3 arc min Level: should. harmonicdrive.net 2026-11-08
  6. Give a camera a second frame with the _optical suffix, with z forward, x right and y down. Level: must. ros.org 2026-11-08“z forward x right y down”
  7. Verify camera assignments after any replug, because indices can change. Level: must. docs.nvidia.com 2026-11-08
  8. Do not assume current maps to torque at gear ratios well above 10; backlash, vibration and friction can no longer be ignored. Level: guide. manipulation.csail.mit.edu 2026-11-08