Design Book › Sensor integration

Sensor integration

What a designer must do

  1. Put all streams on one clock with low jitter, and stamp each sample at its measurement time [1], [2]. Treat the time offset between sensors as a calibration item [3].
  2. Sample at more than twice the signal bandwidth, and filter before you sample [4], [5]. Add up latency from rise time to compute [6].
  3. Calibrate in order. Camera intrinsics come before hand-eye [7]. IMU intrinsics come before camera-IMU [8].
  4. Calibrate on the final assembly. Soldering and reflow change IMU and ToF results [9], [10]. A magnetometer needs a full calibration after you install it [11].
  5. Fix an encoder magnet so it cannot slip, and centre it over the sensor [12], [13]. Do not set it directly in an iron shaft [14].
  6. Do not take load accuracy from motor encoder resolution [15].
  7. Keep the IMU from clipping [16]. On a PX4 flight controller, set any non-default board orientation in the firmware [17]. Do not dead-reckon position from a MEMS IMU alone [18].
  8. Take minimum depth for the resolution you run, then measure it on the unit [19], [20].
  9. Keep every force-torque axis below overload, and count tool weight [21], [22]. Reject all axes when any gauge saturates [23].
  10. Read lidar and ToF range with its reflectivity and light conditions [24], [25]. Do not expect a yaw reference without a magnetometer [26].

Scope: the sensor guidelines. Choosing a part is in Sensors. This chapter covers what happens after you choose: timing, calibration, mounting and limits. It lists the guidelines a designer needs. It leaves out guidelines that repeat them or fit one product only. Each row gives the guideline's current level, statement and check.

Most rows rest on one maker or one tool, so a row names its scope. A row proven on one part does not prove the same point on another. Several IMU and magnetometer rows come from PX4 and ArduPilot, which are flight stacks. The vibration guideline says so directly [27]: not measured on arms, so set your own limit.

"Audit note below" on a row means an audit questioned that guideline. See the Audit notes section near the end of this chapter.

Time sync and rate

Clock and stamps

Clock and stamps
LevelConditionCheckSource
mustTreat time offset between sensors as a calibration item, because it strongly affects fusion. Scope: Visual-inertial systems.measure: estimate camera-IMU time offset and record it[3]
shouldTimestamp with low jitter on a single clock. Scope: Kalibr. Audit note below.measure: histogram of sample period, per sensor, on the common clock[1]
mustInspect IMU timestamp steps; buffered bursts and gaps must be corrected before Kalibr calibration. Scope: Kalibr.measure: plot of IMU dt from the log[28]
shouldStamp each sample with its measurement time, not the time your program read it. Scope: Stereolabs ZED SDK; applies to all sensors. Audit note below.review: driver stamp comes from the sensor or as close to capture as the hardware allows[2]
shouldLearn which instant a camera's stamp marks. This one is the start of readout, which is when the exposure ends. Scope: Stereolabs ZED SDK. Audit note below.calc: stamp correction from the camera's stated stamp instant[29]
shouldOn supported ZED X cameras with SDK 5.5+, use mid-exposure timestamps for instantaneous-sensor fusion; they precede IMAGE by half an exposure. Scope: Stereolabs ZED SDK, 5.9 ms at an 11.7 ms exposure.calc: t_frame = t_exposure_end - exposure / 2[30]
shouldDo not treat the nearest IMU sample as aligned with a frame; it can be a few milliseconds off. Scope: Stereolabs ZED SDK.calc: interpolate IMU data to the frame time[31]
shouldSize an estimator's delay buffer at least as long as the longest sensor delay. Scope: PX4 EKF2. Audit note below.calc: fusion horizon >= max(sensor delays)[32]
shouldDo not treat approximate-time message matching as sensor sync; it pairs by stamps and can drop messages. Scope: ROS message_filters ApproximateTime.measure: dropped and matched set counts and the time spread of each set[33]

A stamp helps only if you know which instant it marks. The ZED SDK stamps an image when the exposure ends. On supported ZED X cameras with SDK 5.5+, the mid-exposure stamp is half an exposure earlier. The guideline's example is 5.9 ms at an 11.7 ms exposure [29], [30].

Hardware sync

Hardware sync
LevelConditionCheckSource
guideAssume multiple cameras are unsynchronised until you wire and configure them. Scope: Intel RealSense D4xx.measure: frame time difference between cameras over a long run[34]
mustDrive the external sync input with the stated pulse at the camera frame rate: 100 us positive, 1.8 V CMOS. Scope: Intel RealSense D4xx.measure: pulse width and level on an oscilloscope[35]
shouldDrive the RealSense HW SYNC input with 1.8 V CMOS levels. Scope: Intel RealSense D4xx.measure: sync level at the camera input[36]
shouldDo not fuse lidar data stamped in internal oscillator mode with other sensors; that clock is a free-running count from turn on. Scope: Ouster lidar. Audit note below.review: lidar timestamp mode is set and recorded[37]
shouldUse IEEE 1588 PTP or an external pulse to sync a lidar with the rest of the robot. Scope: Ouster lidar.measure: PTP lock status and offset during operation[38]

Passive sync wiring suits runs under about 3 m [39]. The OS1 states under 1 ms sync error in PTP mode [40]. The Mid-360 guideline names PTP or GPS as its sync methods [41]. Check that your robot can supply one before you select the lidar.

Rate and latency

Rate and latency
LevelConditionCheckSource
mustSample at more than twice the signal bandwidth to reconstruct it free of aliasing.calc: sample_rate > 2 * signal_bandwidth[4]
mustLow-pass filter a signal before sampling so high frequencies cannot alias.calc: filter attenuation at half the sample rate[5]
shouldDo not read sensor bandwidth from its sample rate, since they are separate properties. Audit note below.review: selection table lists bandwidth and sample rate separately[42]
shouldStart the latency budget with sensor rise time, which falls as bandwidth rises.calc: total latency = sensor rise time + filter delay + transport + compute[6]
shouldInclude IMU filter group delay in the loop budget; it falls as output rate rises. Scope: Bosch BMI270 gyro, group delay at 200 Hz ODR. Value on file: 5.97 ms.calc: phase_lag = 360 * f * group_delay[43]
guideStart from about 1 kHz for commanded torque, with the amplifier's inner loop 10 times faster or more. Scope: Modern Robotics, typical robot arm.calc: inner loop rate >= 10 * torque command rate[44]
guideSample at least 20 times the required closed-loop bandwidth as a starting lower bound. Scope: Secondary citation of Franklin, Powell and Workman. The book was not captured.calc: control_rate >= 20 * closed_loop_bandwidth[45]

In visual-inertial use, run the IMU well above the frame rate [46]. Force-torque rate rows are in the Force-torque section below. The bus adds time too. For serial-bus servos, see TTL serial bus. On EtherCAT, take position samples at equal intervals, since velocity is derived from them [47]. Log the command transport delay for your model [48].

Calibration

The guidelines fix three orderings:

Camera intrinsics

Camera intrinsics
LevelConditionCheckSource
shouldCollect at least 10 good views of the calibration pattern in different positions. Scope: OpenCV calibration.review: calibration log lists at least 10 accepted views[50]
shouldMove the target to the left, right, top and bottom of the view, tilt it, and let it fill the view. Scope: ROS camera_calibration.measure: detected corners cover all image regions[51]
guideAim for reprojection error below 0.2 to 0.5 pixel in offline calibration. Scope: OpenVINS offline calibration guidance.measure: final reprojection error[52]
mustReject a calibration whose radial distortion is not monotonic. Scope: OpenCV camera model.calc: radial distortion function is monotonic over the used image radius[53]
shouldRescale focal length and centre when using a resolution other than the calibrated one; distortion coefficients stay the same. Scope: OpenCV camera model.calc: fx, fy, cx, cy scaled by the resolution ratio[54]
mustEnsure the sensor is in focus before calibrating. Scope: OpenVINS and Kalibr workflow. Audit note below.measure: focus check before data capture[49]

Stereo and multi-camera

Stereo and multi-camera
LevelConditionCheckSource
shouldExpect stereo calibration to shift over time with mechanical stress, vibration and thermal expansion, and plan a correction. Scope: Stereolabs ZED cameras.measure: depth on a known target after thermal and vibration exposure[55]
guideRecalibrate a stereo camera when plane-fit RMS exceeds about 0.2 pixels in the central area. Scope: Intel RealSense D4xx.measure: plane-fit RMS on a flat textured target[56]

Do not compare the two pixel limits. The first is reprojection error in offline calibration [52], and the second is plane-fit RMS on a stereo camera [56]. They measure different things.

Camera-IMU

Camera-IMU
LevelConditionCheckSource
mustMove with at least one translation and two axes of rotation, or camera-IMU parameters are not observable. Scope: Camera-IMU calibration.review: planned trajectory excites all IMU axes[57]
mustAccept a camera-IMU result only if IMU errors and estimated biases stay within their 3-sigma bounds. Scope: OpenVINS and Kalibr workflow.measure: read the error plots in the calibration report[58]
shouldCompare calibrated rotation and translation with hand measurements. Scope: OpenVINS and Kalibr workflow.measure: CAD or ruler values against the calibrated transform[59]

Encoder mounting and magnets

Resolution is not load accuracy

Resolution is not load accuracy
LevelConditionCheckSource
guideChoose encoder resolution at least 2 to 4 times finer than the required position accuracy. Scope: maxon rule of thumb, motor shaft.calc: encoder_step <= required_accuracy / 2[60]
guideFor field-oriented control of a brushless motor, use at least a 500 count-per-turn incremental or an 11-bit absolute encoder, stiffly on the motor side. Scope: maxon rule of thumb.calc: encoder counts per turn >= 500, or absolute bits >= 11[61]
mustDo not take load accuracy from motor encoder resolution: backlash and drive-train elasticity can make the load far worse.calc: load_error >= gear backlash + drive-train windup[15]
guideAssume backlash of 0.3 to 2 degrees for an ordinary gearhead until you have the gearbox figure. Scope: maxon general statement; the upper bound is 2 degrees.measure: load angle hysteresis at a fixed motor angle[62]
shouldTo position a load through backlash, use a low-backlash drive or dual-loop control with a load-side encoder.measure: load position error with direction reversal[63]
shouldCompare the encoder step with the stated backlash before you quote accuracy. Example: 4096 pulses per turn beside 0.25 degrees of backlash. Scope: ROBOTIS XM430-W350.calc: encoder_step = 360 / counts_per_rev, compared with backlash[64]
guideWith no magnet displacement, no zero-programming and a 360° maximum angle, the AS5600 has ±1° system nonlinearity at 12-bit resolution. Scope: ams AS5600, no magnet displacement.measure: encoder angle against a reference over one full turn[65]

Compare the step with the backlash, as the encoder-step guideline asks [64]. The step and the ratio are computed: step is 360 over counts per turn [66].

Resolution is not load accuracy
ServoCounts per turnStep, computedBacklashBacklash over step, computed
XM430-W3504096 [64]0.0879 deg0.25 deg, stated [64]2.8
STS3215-C0184096 [67]0.0879 degabout 1 deg, bench test [68]about 11, at 1 deg

Magnet mounting

Magnet mounting
LevelConditionCheckSource
mustOn a controller's commutation encoder, fix the sense magnet to the rotor and not to a reducer or linkage output. Scope: moteus controllers.review: drawing shows the sense magnet on the rotor[69]
mustFix the magnet rigidly so it cannot slip; adhesive is generally required. Scope: moteus controllers.review: magnet retention method stated on the drawing[12]
mustFor the default onboard moteus encoder, centre it on the sense magnet with about a 2 mm air gap. Scope: moteus controllers.measure: air gap and lateral offset at assembly[70]
shouldIf using a 3D printed encoder mount, make sure it does not warp; PLA and other low temperature filaments warp under load at room temperature. Scope: moteus guidance; applies to any printed encoder mount.measure: air gap after a thermal soak under load[71]
mustMount an on-axis magnet centred over the sensor package. Scope: ams on-axis magnetic position sensors.measure: lateral offset between magnet axis and package centre[13]
shouldFor ams on-axis sensors, air gap changes flux density and linearity; x-y displacement also changes linearity. Scope: ams on-axis magnetic position sensors.calc: tolerance stack for air gap and for lateral offset[72]
mustDo not set the magnet directly in an iron shaft, because the field is weakened substantially. Scope: ams on-axis magnetic position sensors.review: shaft material at the magnet is non-magnetic, or a spacer is present[14]
shouldCheck the stainless grade next to a sense magnet, because some grades are magnetic. Scope: ams on-axis magnetic position sensors.measure: magnet pull test on the actual shaft material[73]
guideFor an AS5048 with an 8 mm magnet displaced 500 µm in x and y, nonlinearity is ±0.7°. Scope: ams AS5048, 8 mm magnet.calc: encoder error budget uses the displaced-magnet figure, not the optimum[74]
shouldKeep the field at the die within the sensor's stated range across the air gap tolerance. Scope: ams AS5048.calc: magnet field at minimum and maximum air gap, 30 to 70 mT[75]
shouldRead the sensor's field diagnostics at run time: the flags indicate a weak or high magnetic field, and the maker recommends monitoring the magnitude value too. Scope: ams AS5048. Audit note below.review: firmware reads the field flags and magnitude and faults on them[76]
shouldStay in the recommended field range on this sensor; signal-to-noise ratio is greater at greater field. Scope: Allegro A1333. Audit note below.calc: field at the air gap is 300 to 1000 G[77]

From -40 to 150°C, the same AS5048 offset gives ±1.2° [78]. If the shaft must be magnetic, put a non-magnetic spacer between it and the magnet [79].

The field ranges are per sensor and in the source's units: 30 to 70 mT on the AS5048 [75], 300 to 1000 G on the A1333 [77]. To reduce misalignment on the A1333, use a smaller air gap and a larger magnet [80]. The 2 mm air gap is for moteus [70]. It is not a general value.

Multi-turn, zero and speed

Multi-turn, zero and speed
LevelConditionCheckSource
guideXM430-W350 Present Position resets to a value within one rotation at power-on, on reboot, and when Position Control Mode is set or torque turns on in it. Scope: ROBOTIS XM430-W350.measure: power cycle at a known multi-turn pose and compare the reported position[81]
shouldA single-turn magnetic sensor keeps a turn count through power-off only if a low-power counter stays supplied. Scope: Allegro A1339. Audit note below.measure: rotate the magnet while the system is off and read the count at wake[82]
shouldBudget dynamic angle error as speed times the sensor's propagation delay, and use a compensated output at speed. Scope: ams AS5047P.calc: DAE = omega * t_delay[83]
shouldPlan one-time zero programming carefully; the burn command works at most 3 times. Scope: ams AS5600.review: assembly sequence writes the zero position only after final mechanical alignment[84]
guideRun this off-axis ring encoder's self-calibration after installation to reach the stated accuracy. Scope: RLS AksIM-2. Value on file: plus or minus 0.004 to 0.020 deg after encoder self-calibration. Audit note below.measure: run self-calibration on the installed encoder[85]

IMU mounting and noise

Axes and mounting

Axes and mounting
LevelConditionCheckSource
mustIMU axes must form a right-handed frame; a driver that receives left-handed data flips the y axis before publishing. Scope: ROS IMU drivers, REP-145 draft.measure: rotate the IMU about each axis by hand and confirm positive gyro output by the right-hand rule[86]
mustAt rest with body z up, the accelerometer z output must read +g in m/s^2, because it reports specific force and not coordinate acceleration. Scope: ROS IMU drivers, REP-145 draft.measure: at rest the upward axis reads +g and the units are m/s^2[87]
shouldIf a PX4 controller cannot sit at the centre of gravity, set EKF2_IMU_POS_X/Y/Z to its offset. Scope: PX4 flight controllers.sim: pitch about the CoG with the offset set; the position estimate stays still[88]
mustOn a PX4 flight controller, configure any non-default board orientation in the firmware. Scope: PX4 flight controllers.measure: tilt the robot in each axis and confirm the estimate tilts the same way[17]
shouldCalibrate misalignment across the full chain from IMU sensing axes to the frame you care about; the factory case alignment is not enough for high performance.measure: calibrate IMU axes against the camera or link frame after final assembly[89]

Vibration

Vibration
LevelConditionCheckSource
shouldCount the isolator as a filter in the control loop: a soft mount adds latency. Scope: PX4 multicopter tuning.measure: phase lag from frame to IMU through the mount at the control bandwidth[90]
shouldReduce the vibration source first; adjust software filters only as a last step. Scope: PX4 airframe guidance.measure: vibration spectrum before and after each mechanical fix[91]
mustKeep the IMU from clipping under vibration; a clipped sensor can cause estimator failure. Scope: PX4 airframe guidance.measure: peak acceleration at the mount stays below the selected full-scale range[16]
guideAs a starting limit from multicopters, treat vibration below 30 m/s/s as normal and above 60 m/s/s as a likely fault. Scope: ArduPilot Copter vibration metric; not measured on arms, so set your own limit.measure: high-pass filtered accelerometer vibration level over a full motion cycle[27]
shouldPick an accelerometer range with headroom for peak vibration; the BMI088 offers up to 24 g for this reason. Scope: Bosch BMI088.calc: peak mount acceleration stays below the selected range[92]

Assembly

Assembly
LevelConditionCheckSource
mustDo not solder the exposed die pad of this IMU to the PCB; package stress changes its performance. Scope: TDK ICM-20948.review: PCB footprint has no solder paste on the die pad[93]
mustRun sensor calibration on the assembled product, because soldering to the board can change the parameters. Scope: CEVA BNO085/BNO086.measure: repeat calibration after the sensor is on the final board[9]
mustCalibrate the D435i IMU yourself: it ships without internal calibration, so idle gyro output is non-zero and accelerometer gravity is off. Scope: Intel RealSense D435i.measure: after calibration, idle gyro is near zero and accelerometer norm is near gravity[94]

Noise budgets

Noise budgets
LevelConditionCheckSource
shouldCompare IMUs by noise density, because a noise sigma depends on the output rate.calc: sigma = noise_density * sqrt(sample_rate)[95]
shouldBudget drift from gyro noise as random walk times the square root of time.calc: sigma_angle = ARW * sqrt(t)[96]
shouldGive in-run bias stability the most weight when selecting an IMU, as it sets the floor on bias knowledge.review: compare in-run bias stability across candidate IMUs at constant temperature[97]
guideTurn-on bias varies between starts, so production calibration cannot remove it; an aided inertial navigation system can estimate it after each startup.measure: bias estimate from a still period at start-up, compared across power cycles[98]
shouldBudget attitude error from uncorrected gyro bias as bias times time. Audit note below.calc: attitude_error = residual_bias * t[99]
shouldBudget temperature-driven bias separately; the bias stability figure excludes it.calc: temperature bias term added to the bias stability term[100]
shouldBudget gyro offset change with temperature from the temperature coefficient. Scope: Bosch BMI088, typical. Value on file: 0.015 deg/s per K.calc: offset_change = TCO * delta_T[101]
mustDo not dead-reckon position from a MEMS IMU alone; one simple INS test passed 150 m of average error after 60 s. Scope: Simple INS on an Xsens Mtx IMU; a MEMS result, not a general limit.sim: propagate measured IMU errors and confirm an aiding sensor bounds position error[18]
shouldDo not feed static noise figures straight into a moving estimator: the model leaves out scale factor errors and temperature bias changes.review: estimator noise is inflated above the static Allan values[102]
shouldCheck gyro acceleration sensitivity on a vibrating robot: oscillating acceleration can shift the gyro bias.measure: gyro bias with the mount vibrating against bias at rest[103]

A worked example. The gyro noise densities that the database holds for parts in Sensors come from sources that do not let us copy them, so the densities below are design assumptions. The results are computed with the noise-density and random-walk formulas [95], [96]. The 200 Hz rate is the starting IMU rate for camera-IMU calibration [104], and the 60 s is the time span of the dead-reckoning test [18].

Noise budgets
Gyro noise density, assumedSigma at 200 Hz, computedAngle sigma after 60 s, computed
5 mdps/rtHz70.7 mdps38.7 mdeg
10 mdps/rtHz141.4 mdps77.5 mdeg

This is the white noise term only. It leaves out bias, scale factor and temperature [102].

Allan log and estimator inputs

Allan log and estimator inputs
LevelConditionCheckSource
mustRecord at least 3 hours of still IMU data for an Allan analysis; longer gives better accuracy. Scope: allan_variance_ros tool.measure: logged static duration >= 3 h[105]
guideKalibr's authors recommend about 15 to 24 hours of stationary IMU data for Allan analysis. Scope: Kalibr IMU noise setup.measure: logged static duration >= 15 h[106]
guideOpenVINS recommends collecting upwards of 20 hours of stationary IMU data. Scope: OpenVINS IMU noise setup.measure: logged static duration >= 20 h[107]
mustCalibrate IMU scale, axis misalignment and nonlinearity before camera-IMU calibration, or use a model that estimates them.review: IMU intrinsics are applied, or the camera-IMU tool's IMU model estimates them[8]

The three record lengths differ, and the database does not settle them. Each comes from its own tool.

Camera and depth limits

Lens and shutter

Lens and shutter
LevelConditionCheckSource
shouldDo not rely on 2 to 3 percent distortion being harmless when you measure from the image; that figure holds only without measurement algorithms.calc: pixel error at the image edge from the distortion percentage[108]
shouldEstimate rolling shutter skew from row read-out time, row count and image speed. Audit note below.calc: skew_px = row_time * rows * image_speed_px_per_s[109]
mustDo not expect shorter exposure to cure rolling shutter skew; it only reduces blur.review: skew budget does not credit exposure time[110]
shouldChoose a global shutter where things move fast in the image, and pair it with short exposure.calc: image speed against the skew and blur budgets[111]
mustDo not use a short strobe with a rolling shutter unless the flash covers all rows, or you get a bright band.measure: image under the strobe shows even exposure across all rows[112]
shouldKeep exposure short enough that the target moves less than about one pixel during it.calc: image_speed_px_per_s * exposure_s <= 1 px[113]
mustDo not expect global shutter to remove motion blur; only exposure does that.calc: blur budget applies to both shutter types[114]
mustModel rolling-shutter effects in direct visual-inertial odometry; neglecting them degrades accuracy and robustness. Scope: Direct visual-inertial odometry.sim: VIO error with and without the rolling shutter model on recorded data[115]
shouldCheck depth imager shutter before you select a stereo module for a moving arm; modules in one family differ. Scope: Intel RealSense D400 family, rolling shutter models.review: stereo module's depth imager shutter type against the expected motion[116]

Stereo depth

Stereo depth
LevelConditionCheckSource
shouldFor RealSense D400 cameras, depth error grows with distance squared; work as close as MinZ allows. Scope: Intel RealSense D400.calc: depth_rms(z) from baseline, focal length and subpixel figure, at maximum working range[117]
shouldUse the ZED accuracy range, about 1 percent near and about 9 percent at maximum range, as the working budget. Scope: Stereolabs ZED cameras.measure: depth on a known target at near and far distance[118]
shouldUse the maker's per-range error: below 2 percent under 4 m, then below 4 percent to 7 m and below 6 percent to 10.5 m. Scope: Luxonis OAK 800P, 75 mm baseline, median device.measure: depth on a known target at each range band[119]
shouldRead the D400 2 percent figure only with its scope: within 80 percent of the field and within the stated distance. Scope: Intel RealSense D400 family, 80 percent ROI.review: working range and image region are inside the spec conditions[120]
shouldTrade baseline with care: a wider baseline improves depth accuracy but raises the minimum distance. Scope: Luxonis OAK; applies to stereo in general.calc: min_distance = focal_px * baseline / max_disparity[121]
shouldCalculate minimum stereo distance from focal length, baseline and maximum disparity. Scope: Luxonis OAK; applies to stereo in general. Audit note below.calc: min_distance = focal_px * baseline / max_disparity[122]
shouldIf D400 MinZ prevents getting closer, reduce depth X-resolution to reduce MinZ. Scope: Intel RealSense D400.calc: MinZ scales in proportion to X-resolution[123]
mustCheck exposure first when depth quality is poor, as both under and over exposure hurt. Scope: Intel RealSense D400.measure: histogram of left and right images in the task scene[124]
shouldSet stereo depth gain low and adjust exposure first; gain adds noise that degrades depth. Scope: Intel RealSense D400.measure: plane-fit RMS at each gain setting[125]
shouldProvide texture for stereo: passive matching quality depends on visually distinct points. Scope: Stereo depth. Audit note below.measure: fill rate on the real surfaces in the task[126]
shouldTest white walls, plain surfaces and reflective areas, because they produce outliers in stereo depth. Scope: Stereolabs ZED cameras; applies to stereo in general. Audit note below.measure: outlier rate on the worst surface in the task[127]
shouldDesign to the camera's ideal depth range, not its maximum. ZED Mini: 15 m maximum and 9 m ideal. Scope: Stereolabs ZED Mini.review: working range is inside the ideal band[128]

The ZED advice on gain differs [129]. See Where the sources disagree.

Minimum depth on the D435

Two maker documents give different values for the same D435 mode. The database keeps both and does not settle them. Measure your unit.

Minimum depth on the D435
DocumentD435 at 848x480SourceLevelCheck
Datasheet 337029-013195 mm[19]shouldmeasure: closest valid depth on the real unit at the used resolution
Tuning paperabout 16.8 cm[20]guidemeasure: closest valid depth on the real unit

ToF cameras, cables and power

ToF cameras, cables and power
LevelConditionCheckSource
mustTreat depth value 0 as invalid data, not as a distance. Scope: Azure Kinect DK.review: consumers reject depth value 0[130]
guideSaturated IR pixels lose phase information and are invalidated in Azure Kinect DK depth. Scope: Azure Kinect DK.measure: invalid pixel count on shiny parts[131]
shouldExpect invalid depth at corners with a ToF camera, as light from more than one object reaches a pixel. Scope: Azure Kinect DK.measure: invalid pixels and depth error in the corners of the workspace[132]
shouldRead ToF operating range with its conditions: reflectivity, ambient IR, and error limits. Scope: Azure Kinect DK. Value on file: random error with a standard deviation of 17 mm or less, within the reflectivity and ambient-IR conditions the maker states.review: the task's reflectivity and ambient light are inside the stated conditions[133]
shouldCheck ToF range for each mode. Narrow unbinned runs 0.5 to 3.86 m. Scope: Azure Kinect DK, NFOV unbinned.calc: working distance inside the range for the selected mode[134]
shouldUse short high-quality USB3 cables for depth cameras, preferably under 1 m, and active repeaters beyond 2 m. Scope: Intel RealSense D4xx.measure: dropped frames at full data rate on the final harness[135]
shouldBudget power for each depth camera; a camera with the projector on can reach or exceed 2 W. Scope: Intel RealSense D4xx.calc: sum of camera power against the port or hub limit[136]

Hand-eye

Rows rest on OpenCV and MoveIt, plus tips listed in one README.

Hand-eye
LevelConditionCheckSource
mustOpenCV hand-eye calibration needs at least 2 motions with non-parallel rotation axes, so at least 3 poses; many more are strongly recommended. Scope: OpenCV calibrateHandEye.calc: rotation axes of the motion set are not all parallel[137]
guideExpect a first result after 5 samples; quality improves with more and typically plateaus at about 12 or 15. Scope: MoveIt Calibration. Audit note below.measure: calibration change as samples are added[138]
mustRotate between each pair of poses, and about more than one axis. Scope: MoveIt Calibration.calc: rotation magnitude and axes between poses[139]
mustKeep the target stationary in the robot base frame during eye-in-hand calibration. Scope: MoveIt Calibration, eye-in-hand.measure: target has not moved at the end of the run[140]
mustCalibrate camera intrinsics before hand-eye calibration. Scope: MoveIt Calibration.review: intrinsic calibration result is on file before hand-eye capture[7]
shouldMaximise rotation between poses. Scope: Tips from Tsai and Lenz, as cited in the easy_handeye README.calc: rotation angle between pose pairs[141]
shouldKeep the target close to the camera during hand-eye calibration. Scope: Tips from Tsai and Lenz, as cited in the easy_handeye README.calc: camera to target distance in the pose plan[142]
shouldMinimise translation between poses during hand-eye calibration. Scope: Tips from Tsai and Lenz, as cited in the easy_handeye README.calc: translation between pose pairs[143]
shouldCalibrate the robot if necessary or applicable. Scope: Tip listed in the easy_handeye README. Audit note below.measure: repeat hand-eye after robot kinematic calibration[144]
guidePrefer a method that solves rotation and translation together when the data is noisy; the authors' comparison of three methods found it seems the most robust. Scope: Horaud and Dornaika, Hand-Eye Calibration. Audit note below.sim: compare solvers on the recorded pose set[145]
mustKeep the target flat, or it cannot be localised reliably. Scope: MoveIt hand-eye calibration target.measure: flatness of the mounted target[146]

The Tsai and Lenz tips are second hand: the easy_handeye README lists them, and the original paper was not captured. Do not carry them over to camera-IMU work. Hand-eye asks you to minimise translation [143], and camera-IMU needs at least one translation [57].

Force-torque

Rows are ATI F/T transducers where no other scope is shown. Part data is in Sensors.

Force-torque
LevelConditionCheckSource
mustKeep every load axis below the sensor's single-axis overload value, as exceeding it can cause irreparable damage.calc: peak force and torque on each axis, including impact, against the overload table[21]
mustProtect small force-torque sensors during assembly; bolting them up can overload them. Scope: ATI Nano and Mini transducers.review: installation procedure limits load on the sensor while fasteners are tightened[147]
mustLook up overload values per model and per axis. A Nano17 Titanium, for example, is rated to 160 N in x-y and 310 N in z. Scope: ATI Nano17 Titanium.calc: load case against the table for the exact model[148]
mustCheck combined loads, because a complex load can saturate the gauges while each axis is below its rating.calc: run the maker's complex-loading check for the worst load combination[149]
mustMonitor for saturation and reject all axis readings when any gauge saturates.review: controller stops force control on the saturation flag[23]
mustCount tool weight against overload even after biasing it out of the readings. Scope: ATI Net F/T.calc: tool_load + task_load < overload, using unbiased data[22]
shouldTake a bias or reference reading at the current temperature just before applying the load you want to measure.measure: unloaded reading is recorded at working temperature before each measurement[150]
shouldDo not assume compensation removes zero drift: temperature compensation covers gain, and drift varies per unit.measure: unloaded output against temperature for the unit you use[151]
shouldAdd temperature gain error to rated room-temperature accuracy for an overall estimate, if unloaded and loaded readings share a temperature; ATI non-Gamma transducers show 0.5% typical gain error at ±15°C from 22°C. Scope: ATI non-Gamma transducers.calc: overall_accuracy = rated_accuracy + gain_error(delta_T)[152]
guideATI non-Gamma transducers show 5% typical gain error at ±50°C from 22°C, bounded by their operational limits. Scope: ATI non-Gamma transducers.calc: gain_error from the maker's temperature table[153]
mustRecalculate range, resolution and accuracy when you move the point of origin; published figures hold at the factory origin only.calc: load at the custom tool point transformed to the sensor origin[154]
mustRestrain the sensor cable close to its connector so repetitive motion does not reach the connector.review: harness drawing shows a clamp next to the connector[155]
shouldRebias or compensate during long runs; the authors report that temperature drift can be considerable. Scope: Six-axis strain gauge sensors on the iCub robot.measure: unloaded output at the end of a long run[156]
shouldUse a force-torque sensor's anti-alias filter, or provide your own where the system has none. Scope: ATI Mux transducers; ATI states the filter is absent in 9105-TIF and TWE. Audit note below.review: anti-alias filter is listed in the signal chain[157]
shouldRead a force-torque filter's group delay before closing a force loop on it. Scope: ATI Net F/T.calc: group delay at the cutoff against the force loop period[158]

The temperature figures are deviations. The first figure reads plus or minus 15 C [152], and the second reads plus or minus 50 C [153]. Both are measured from 22 C. Do not read them as a wrist at 15 C or 50 C.

Lidar and ToF

Lidar

Lidar
LevelConditionCheckSource
mustDo not plan on lidar data inside the blind zone. Livox Mid-360: 0.1 m. Scope: Livox Mid-360.calc: nearest object or robot part against the blind zone[159]
mustTreat returns from just outside the blind zone as reference only: precision is not guaranteed there. Scope: Livox Mid-360, 0.1 to 0.2 m.review: consumers do not use returns from 0.1 to 0.2 m for safety or fine geometry[160]
shouldCheck minimum range for the exact model and firmware. Ouster OS1: 0.5 m in this datasheet. Scope: Ouster OS1 rev7, firmware before 3.1.calc: nearest object against the model's minimum range[161]
guideTest any zero-blind-range claim on your own unit. Scope: Hesai XT32, maker claim.measure: valid returns from a target moved to the enclosure[162]
guideAt 100 klx, 1024×10 Hz and over 90% detection probability, OS1 rev7 ranges 90 m at 10% and 170 m at 80% Lambertian reflectivity. Scope: Ouster OS1 rev7.calc: required range against the 10 percent reflectivity figure[163]
shouldRead 2D scanner range by reflectance: this scanner drops from 10 m on white to 4 m at 10 percent reflectance. Scope: Hokuyo UST-10LX. Audit note below.calc: required range against the low-reflectance figure[164]
mustCheck surrounding light limit, and avoid direct sun on this scanner. Scope: Hokuyo UST-10/20LX. Value on file: less than 80000 lx.measure: scan quality in the brightest expected light[165]
shouldRead lidar range with sunlight and reflectivity conditions. Scope: Livox Mid-360. Value on file: 40 m at 10 percent reflectivity. Audit note below.calc: required range against the 10 percent figure[24]

Near limits are per unit, so these rows do not disagree. Three units give three figures: 0.1 m [159], 0.5 m [161] and a maker claim of zero [162]. Test the claim on your own unit.

ToF range sensors

All numbers below are for the ST VL53L1X, except the temperature-recalibration row, which is the VL53L0X [166].

ToF range sensors
LevelConditionCheckSource
shouldSize a ToF range sensor for bright light, not dark. VL53L1X long mode falls from 360 cm in the dark to 73 cm in strong ambient. Scope: ST VL53L1X, long mode, white target, 100 ms budget. Audit note below.measure: maximum valid range in the brightest expected light on the real target[167]
shouldChoose short distance mode in bright light: range holds at about 135 cm. Scope: ST VL53L1X, short mode, white target, 100 ms budget.measure: maximum valid range in short mode in bright light[168]
shouldCheck range at the target's reflectance and the expected light together; grey targets at 17 percent fall further. Scope: ST VL53L1X, long mode. Value on file: 68 cm in long mode.measure: maximum range on the darkest target in the brightest light[25]
mustRepeat RefSPAD and crosstalk calibration whenever you add a cover glass over the sensor. Scope: ST VL53L1X.measure: range error on a known target with the glass fitted[169]
mustRun offset calibration on every design, because reflow affects it. Scope: ST VL53L1X.measure: range error on a known target after assembly[10]
mustLoad stored calibration data into the sensor at each start. Scope: ST VL53L1X.review: start-up code loads the stored calibration[170]
mustSelect the timing budget for the distance mode; 20 ms works only in short mode, and 33 ms is the minimum for all modes. Scope: ST VL53L1X.review: timing budget and distance mode pair is valid[171]
shouldCheck that the target fills the sensor's field of view, because the published range tables assume that. Scope: ST VL53L1X.calc: target size at range against the 27 degree field of view[172]
mustRepeat temperature calibration when temperature moves more than 8 C from the last calibration. Scope: ST VL53L0X.measure: on-board or host temperature monitor triggers recalibration[166]

Maximum distance, white target, 100 ms budget, from the ST range table [167]:

ToF range sensors
ModeDark, cmStrong ambient, cmSource
Shortnot shownabout 135[167], [168]
Mediumnot shownnot shown[167]
Long36073[167]

Office light is near 5 kcps/SPAD, and 200 kcps/SPAD is direct sun through a window [173]. The target must fill the field of view [172]. With the 27 deg field and the footprint formula [174], the footprint is 173 cm at 360 cm and 35 cm at 73 cm. Both are computed.

Sensor fusion

Orientation outputs

Orientation outputs
LevelConditionCheckSource
mustWithout a magnetometer, expect only roll and pitch to be referenced; yaw is arbitrary. Scope: ROS IMU drivers, REP-145 draft.review: no function depends on an absolute yaw from a 6-axis fusion output[26]
shouldPlan for offsets that change during a run: background calibration in fusion modes cannot be disabled. Scope: Bosch BNO055.measure: orientation repeatability across runs with calibration running[175]
mustTest pitch and roll drift under vibration or constant motion, because the fusion ignores the accelerometer during motion. Scope: Bosch BNO055.measure: pitch and roll against a reference through a long motion cycle[176]
shouldTest any sustained acceleration case, because the fusion can read it as gravity. Scope: Bosch BNO055. Audit note below.measure: orientation error during the longest sustained acceleration in the task[177]
shouldValidate built-in fusion on robot motion; this algorithm was tuned for human motion. Scope: Bosch BNO055.measure: compare fused orientation with joint-encoder orientation on the arm[178]
shouldUse the game rotation vector where a heading jump would hurt and slow yaw drift is acceptable; it does not use the magnetometer. Scope: CEVA BNO08x.measure: yaw drift over the task duration[179]
guideFor simulated BNO08x motion with nominal calibration, gaming rotation-vector dynamic heading drift is 0.5°/min. Scope: CEVA BNO08x, nominal calibration.calc: heading_drift = 0.5 deg/min * task_minutes[180]
guideFor simulated BNO08x motion with nominal calibration, rotation-vector error is 3.5° dynamic and 2.0° static. Scope: CEVA BNO08x, nominal calibration.calc: orientation budget uses the composite sensor error[181]
shouldAvoid the accelerometer-and-magnetometer output on a fast or magnetically noisy robot: it responds slowly and errs in varying fields. Scope: CEVA BNO08x.measure: orientation error during fast motion with motors running[182]

Magnetic heading

Magnetic heading
LevelConditionCheckSource
shouldUse a magnetometer fusion mode only where the surrounding field is stable; the maker warns that fields around electronics are usually much larger than the Earth's. Scope: Bosch BNO055. Audit note below.measure: heading change while motors run at full current[183]
shouldMount the magnetometer away from power distribution and other magnetic sources. Scope: ArduPilot Copter guidance.measure: magnetic field reading at no load and full motor current[184]
shouldKeep DC power leads short and twist them to limit field at the magnetometer. Scope: ArduPilot Copter guidance.review: harness drawing shows short twisted power leads near the sensor[185]
mustDo a full hard-iron and soft-iron calibration after installing the sensor on the robot, and after large configuration changes. Scope: PX4 compass calibration.measure: heading error at several headings after calibration[11]
shouldDo not treat offset-only calibration as complete; it corrects hard iron only. Scope: PX4 compass calibration.review: the calibration type used is recorded with the result[186]
shouldCalibrate well away from strong magnetic objects, or their effect gets saved into the correction. Scope: CEVA BNO085/BNO086.review: calibration site is several feet from magnetic objects[187]

Estimators

Estimators
LevelConditionCheckSource
shouldFuse a second orientation source only if its covariance is honest; if not, use only the more accurate source or fuse it differentially. Scope: ROS robot_localization.review: covariance of each orientation source checked against measured error[188]
mustTransform IMU data to ENU before robot_localization; the node does not work with NED data. Scope: ROS robot_localization, version captured.measure: yaw increases counter-clockwise from above[189]

Position needs an aiding sensor [18]. Size the estimator's delay buffer for the slowest sensor [32].

Where the sources disagree

The guidelines keep both sides. The database does not settle these. Do not average them.

Where the sources disagree
TopicOne sourceOther sourceWhat the guideline saysSources
D435 MinZ at 848x480Datasheet: 195 mmTuning paper: about 16.8 cmMeasure on the unit[19], [20]
Blur: gain or exposureZED: manual mode, raise gain before exposureD400: set gain low, adjust exposure firstEach names its own camera family. The D400 advice is about depth quality [125][129], [125]
Allan record lengthallan_variance_ros: at least 3 hKalibr: about 15 to 24 h. OpenVINS: upwards of 20 hEach is tied to its own tool[105], [106], [107]
Noise inflationKalibr: 10 times or more, lowest-cost sensorsOpenVINS: about 10 to 20 timesTest which factor works[190], [191]
Camera-IMU ratesKalibr: about 20 Hz camera, 200 Hz IMUOpenVINS: 200 to 500 Hz with 20 to 30 Hz videoLog the actual rates[104], [46]
Force-torque biasATI: bias just before the loadiCub authors: not practical on a floating basePlan an in situ method there[150], [192]
VL53L1X long rangeST table: 360 cm, dark, white, 100 ms budgetAdafruit breakout: 4000 mm, claim, no condition on the part pageMeasure on the real target[167]
VL53L1X short modeST table: about 135 cm in strong ambient lightST text: about 1.3 mBoth come from ST. The guidelines do not reconcile them[168], [193]

Audit notes

An audit read these guidelines against their sources. The database keeps the guideline and the note side by side, and it does not settle them. Level and statement in the tables above are the current ones. The note text below is from the database. Where a note says must is over-strong, the guideline is now rated should. Bold marks wording in the guideline that the note still disputes.

Audit notes
GuidelineLevelAudit noteWhere it stands
Ensure the sensor is in focus before calibrating. [49]mustSource says only 'Ensure that your sensor is in focus'; 'then keep it' is not in the source.Rated must. The statement no longer says "then keep it".
Estimate rolling shutter skew from row read-out time, row count and image speed. [109]shouldSource only explains what sets the row time offset; no requirement or failure from not estimating skew, so must is over-strong.Rated should. The statement is still an instruction.
Calculate minimum stereo distance from focal length, baseline and maximum disparity. [122]shouldSource only gives the formula as a worked OAK-D example; no requirement or failure, so must is over-strong.Rated should.
Provide texture for stereo: passive matching quality depends on visually distinct points. [126]shouldSource says texture improves accuracy significantly and a projector is useful; it does not require texture, so must is over-strong.Rated should. The statement still says to provide texture.
Test white walls, plain surfaces and reflective areas, because they produce outliers in stereo depth. [127]shouldSource says accuracy 'may also be reduced' on these surfaces and asks for no test, so must is over-strong.Rated should. The statement still says to test, and says these surfaces produce outliers.
A single-turn magnetic sensor keeps a turn count through power-off only if a low-power counter stays supplied. [82]shouldSource says key-off sensors are 'often' powered from the battery; it does not state 'only if' or require anything for must.Rated should. The statement still says "only if".
Read the sensor's field diagnostics at run time: the flags indicate a weak or high magnetic field, and the maker recommends monitoring the magnitude value too. [76]shouldSource says the flags indicate a weak or high field and recommends monitoring magnitude; it never says they show a mechanical problem.Statement no longer claims a mechanical problem.
Stay in the recommended field range on this sensor; signal-to-noise ratio is greater at greater field. [77]shouldSource does not tie the 300 to 1000 G recommended range to best noise; it says signal-to-noise is greater at greater field, in a sentence cut off in the capture.The statement keeps the two points apart. The capture is cut off, so the full source sentence is not on file.
Run this off-axis ring encoder's self-calibration after installation to reach the stated accuracy. [85]guideCapture never mentions a hollow joint; it describes an off-axis encoder for limited installation space.Statement does not name a hollow joint.
Use a magnetometer fusion mode only where the surrounding field is stable; the maker warns that fields around electronics are usually much larger than the Earth's. [183]shouldSource says compass accuracy depends on field stability; it does not restrict use to stable fields, so 'only where' and must go too far.Rated should. The statement still says "only where".
Test any sustained acceleration case, because the fusion can read it as gravity. [177]shouldThe maker's own wording is 'it is recommended to test', so must is over-strong.Rated should.
Expect a first result after 5 samples; quality improves with more and typically plateaus at about 12 or 15. [138]guideSource says the plateau comes after about 12 or 15 samples; the record drops 15.Statement now says 12 or 15. The value field still holds 12.
Calibrate the robot if necessary or applicable. [144]shouldSource is the bare tip 'Calibrate the robot if necessary / applicable'; the reason about kinematic error is not in it.Statement is now the bare tip. The topic and the check still name robot accuracy and kinematic calibration.
Prefer a method that solves rotation and translation together when the data is noisy; the authors' comparison of three methods found it seems the most robust. [145]guideSource is not a survey: the authors compare their two methods with Tsai and Lenz, and say the method 'seems' most robust.Statement now says the authors' comparison of three methods.
Read 2D scanner range by reflectance: this scanner drops from 10 m on white to 4 m at 10 percent reflectance. [164]shouldSource only lists range for two reflectances; no requirement or failure, so must is over-strong.Rated should.
Read lidar range with sunlight and reflectivity conditions. [24]shouldCapture never calls the Mid-360 solid-state; the word does not appear in it.Statement does not say solid-state. The Mid-360 subcategory and the topic of the lidar-sync guideline still do [41].
Budget attitude error from uncorrected gyro bias as bias times time. [99]shouldSource gives only the linear error formula and calls bias compensation trivial; no requirement or stated failure, so must is over-strong.Rated should.
Use a force-torque sensor's anti-alias filter, or provide your own where the system has none. [157]shouldSource says the filter prevents aliasing and is absent on some models; 'provide your own' and must are not in it.Rated should. The statement still says "provide your own".
Do not read sensor bandwidth from its sample rate, since they are separate properties. [42]shouldSource defines sample rate and bandwidth as different things; no requirement or failure, so must is over-strong.Rated should.
Timestamp with low jitter on a single clock. [1]shouldSource lists this under 'Tips' with no stated failure, alongside tips this file rates should, so must is over-strong.Rated should.
Stamp each sample with its measurement time, not the time your program read it. [2]shouldSource describes what the ZED SDK IMU stamp means; it does not require this of all sensors, so the general must goes further than the source.Rated should. The scope field still says "applies to all sensors".
Learn which instant a camera's stamp marks. This one is the start of readout, which is when the exposure ends. [29]shouldSource describes when the ZED stamp is taken and only says images are compared with light collected earlier; no requirement or failure for must.Rated should.
Do not fuse lidar data stamped in internal oscillator mode with other sensors; that clock is a free-running count from turn on. [37]shouldSource only describes the internal oscillator mode; it does not forbid fusing that data or state a failure, so must goes further than the source.Rated should. The statement still says "do not fuse".
Size an estimator's delay buffer at least as long as the longest sensor delay. [32]shouldSource wording is 'should be at least as large' with no stated failure, so must is over-strong.Rated should.
Size a ToF range sensor for bright light, not dark. VL53L1X long mode falls from 360 cm in the dark to 73 cm in strong ambient. [167]shouldFigures are right (360 dark, 73 strong ambient), but the source only tabulates them and requires no bright-light sizing, so must is over-strong.Rated should. The statement still says to size for bright light.

Where the database is thin

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. Timestamp with low jitter on a single clock. Level: should. github.com 2026-11-09Review note: Source lists this under 'Tips' with no stated failure, alongside tips this file rates should, so must is over-strong.
  2. Stamp each sample with its measurement time, not the time your program read it. Level: should. stereolabs.com 2026-11-09Review note: Source describes what the ZED SDK IMU stamp means; it does not require this of all sensors, so the general must goes further than the source.
  3. Treat time offset between sensors as a calibration item, because it strongly affects fusion. Level: must. arxiv.org 2026-11-09
  4. Sample at more than twice the signal bandwidth to reconstruct it free of aliasing. Level: must. en.wikipedia.org 2026-11-09“the sampling rate must be at least twice the signal's bandwidth”
  5. Low-pass filter a signal before sampling so high frequencies cannot alias. Level: must. en.wikipedia.org 2026-11-09“The type of filter required is a lowpass filter, and in this application it is called an anti-aliasing filter.”
  6. Start the latency budget with sensor rise time, which falls as bandwidth rises. Level: should. vectornav.com 2026-11-09
  7. Calibrate camera intrinsics before hand-eye calibration. Level: must. moveit.github.io 2026-11-09“with good intrinsic calibration parameters and an accurate coordinate frame”
  8. Calibrate IMU scale, axis misalignment and nonlinearity before camera-IMU calibration, or use a model that estimates them. Level: must. github.com 2026-11-09
  9. Run sensor calibration on the assembled product, because soldering to the board can change the parameters. Level: must. ceva-ip.com 2026-11-09
  10. Run offset calibration on every design, because reflow affects it. Level: must. pololu.com 2026-11-09
  11. Do a full hard-iron and soft-iron calibration after installing the sensor on the robot, and after large configuration changes. Level: must. docs.px4.io 2026-11-09“This calibration is required after installing the autopilot on an airframe for the first time”
  12. Fix the magnet rigidly so it cannot slip; adhesive is generally required. Level: must. mjbots.github.io 2026-11-09“The fastening must be rigid, allowing no slip.”
  13. Mount an on-axis magnet centred over the sensor package. Level: must. dtsheet.com 2026-11-09“the magnet must be mounted centred over the IC package”
  14. Do not set the magnet directly in an iron shaft, because the field is weakened substantially. Level: must. dtsheet.com 2026-11-09“The magnet is weakened substantially.”
  15. Do not take load accuracy from motor encoder resolution: backlash and drive-train elasticity can make the load far worse. Level: must. support.maxongroup.com 2026-11-09
  16. Keep the IMU from clipping under vibration; a clipped sensor can cause estimator failure. Level: must. docs.px4.io 2026-11-09“In extreme cases vibration may lead to sensor clipping/failures, possibly resulting in estimation failures and fly-aways.”
  17. On a PX4 flight controller, configure any non-default board orientation in the firmware. Level: must. docs.px4.io 2026-11-09“If the board or any external compasses are mounted in any other orientation then you will need configure this in the firmware.”
  18. Do not dead-reckon position from a MEMS IMU alone; one simple INS test passed 150 m of average error after 60 s. 150 m Level: must. cl.cam.ac.uk 2026-11-09
  19. Take minimum depth from the datasheet table for the resolution you run. D435 at 848x480: 195 mm in this datasheet. 195 mm Level: should. docs.clearpathrobotics.com 2026-11-08
  20. Intel's tuning paper gives D435 MinZ as about 16.8 cm at 848×480. 16.8 cm Level: guide. dev.realsenseai.com 2026-11-09
  21. Keep every load axis below the sensor's single-axis overload value, as exceeding it can cause irreparable damage. Level: must. ati-ia.com 2026-11-09
  22. Count tool weight against overload even after biasing it out of the readings. Level: must. ati-ia.com 2026-11-08
  23. Monitor for saturation and reject all axis readings when any gauge saturates. Level: must. ati-ia.com 2026-11-09
  24. Read lidar range with sunlight and reflectivity conditions. 40 m Level: should. livoxtech.com 2026-11-08Review note: Capture never calls the Mid-360 solid-state; the word does not appear in it.
  25. Check range at the target's reflectance and the expected light together; grey targets at 17 percent fall further. 68 cm Level: should. pololu.com 2026-11-09
  26. Without a magnetometer, expect only roll and pitch to be referenced; yaw is arbitrary. Level: must. ros.org 2026-11-09“the world frame will only be aligned in roll and pitch (due to gravity), while the orientation around the z axis can be arbitrary”
  27. As a starting limit from multicopters, treat vibration below 30 m/s/s as normal and above 60 m/s/s as a likely fault. 30 m/s/s Level: guide. ardupilot.org 2026-11-09“Vibration levels below 30m/s/s are normally acceptable. Levels above 30m/s/s may have problems and levels above 60m/s/s nearly always have problems”
  28. Inspect IMU timestamp steps; buffered bursts and gaps must be corrected before Kalibr calibration. Level: must. github.com 2026-11-09
  29. Learn which instant a camera's stamp marks. This one is the start of readout, which is when the exposure ends. Level: should. stereolabs.com 2026-11-09Review note: Source describes when the ZED stamp is taken and only says images are compared with light collected earlier; no requirement or failure for must.
  30. On supported ZED X cameras with SDK 5.5+, use mid-exposure timestamps for instantaneous-sensor fusion; they precede IMAGE by half an exposure. 5.9 ms Level: should. stereolabs.com 2026-11-09
  31. Do not treat the nearest IMU sample as aligned with a frame; it can be a few milliseconds off. Level: should. stereolabs.com 2026-11-09
  32. Size an estimator's delay buffer at least as long as the longest sensor delay. Level: should. docs.px4.io 2026-11-09“This value should be at least as large as the longest delay”Review note: Source wording is 'should be at least as large' with no stated failure, so must is over-strong.
  33. Do not treat approximate-time message matching as sensor sync; it pairs by stamps and can drop messages. Level: should. wiki.ros.org 2026-11-09“Some messages can be dropped.”
  34. Assume multiple cameras are unsynchronised until you wire and configure them. Level: guide. dev.realsenseai.com 2026-11-09
  35. Drive the external sync input with the stated pulse at the camera frame rate: 100 us positive, 1.8 V CMOS. 100 microsecond Level: must. dev.realsenseai.com 2026-11-09
  36. Drive the RealSense HW SYNC input with 1.8 V CMOS levels. 1.8 V Level: should. dev.realsenseai.com 2026-11-09
  37. Do not fuse lidar data stamped in internal oscillator mode with other sensors; that clock is a free-running count from turn on. Level: should. docs.ouster.com 2026-11-09Review note: Source only describes the internal oscillator mode; it does not forbid fusing that data or state a failure, so must goes further than the source.
  38. Use IEEE 1588 PTP or an external pulse to sync a lidar with the rest of the robot. Level: should. docs.ouster.com 2026-11-09
  39. Use passive sync wiring only under about 3 m, and shielded twisted pair where you can. 3 meters Level: should. dev.realsenseai.com 2026-11-09
  40. Budget under 1 ms sync error for this PTP mode, and verify it if your time budget is tighter. 1 ms Level: should. cdn.shopify.com 2026-11-09
  41. Select lidars with a sync method your robot can supply: PTP or GPS here. Level: should. livoxtech.com 2026-11-08
  42. Do not read sensor bandwidth from its sample rate, since they are separate properties. Level: should. vectornav.com 2026-11-09Review note: Source defines sample rate and bandwidth as different things; no requirement or failure, so must is over-strong.
  43. Include IMU filter group delay in the loop budget; it falls as output rate rises. 5.97 ms Level: should. bosch-sensortec.com 2026-11-08
  44. Start from about 1 kHz for commanded torque, with the amplifier's inner loop 10 times faster or more. 1000 times per second Level: guide. hades.mech.northwestern.edu 2026-11-08
  45. Sample at least 20 times the required closed-loop bandwidth as a starting lower bound. 20 times Level: guide. arxiv.org 2026-11-09
  46. Run IMU rate well above frame rate in visual-inertial use, such as 200 to 500 Hz with 20 to 30 Hz video. 200 hz Level: guide. docs.openvins.com 2026-11-09“Publish your inertial reading at a reasonable rate (200-500hz)”
  47. 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
  48. Estimate the serial command transport delay before fitting, and align command times to the measured rollout. Level: should. raw.githubusercontent.com 2026-11-09“Estimate and log the serial command transport delay ($t_1 - t_0$).”
  49. Ensure the sensor is in focus before calibrating. Level: must. docs.openvins.com 2026-11-09“Ensure that your sensor is in focus”Review note: Source says only 'Ensure that your sensor is in focus'; 'then keep it' is not in the source.
  50. Collect at least 10 good views of the calibration pattern in different positions. 10 snapshots Level: should. docs.opencv.org 2026-11-09“you will probably need at least 10 good snapshots of the input pattern in different positions”
  51. Move the target to the left, right, top and bottom of the view, tilt it, and let it fill the view. Level: should. wiki.ros.org 2026-11-09“checkerboard on the camera's left, right, top and bottom of field of view”
  52. Aim for reprojection error below 0.2 to 0.5 pixel in offline calibration. 0.5 pixel Level: guide. docs.openvins.com 2026-11-09“a good calibration having less than < 0.2-0.5 pixel reprojection errors”
  53. Reject a calibration whose radial distortion is not monotonic. Level: must. docs.opencv.org 2026-11-09“Radial distortion is always monotonic for real lenses, and if the estimator produces a non-monotonic result, this should be considered a calibration failure.”
  54. Rescale focal length and centre when using a resolution other than the calibrated one; distortion coefficients stay the same. Level: should. docs.opencv.org 2026-11-09“these should be scaled along with the current resolution from the calibrated resolution”
  55. Expect stereo calibration to shift over time with mechanical stress, vibration and thermal expansion, and plan a correction. Level: should. stereolabs.com 2026-11-09
  56. Recalibrate a stereo camera when plane-fit RMS exceeds about 0.2 pixels in the central area. 0.2 pixels Level: guide. dev.realsenseai.com 2026-11-09
  57. Move with at least one translation and two axes of rotation, or camera-IMU parameters are not observable. Level: must. docs.openvins.com 2026-11-09“at least one translational motion along with two degrees of orientation change”
  58. Accept a camera-IMU result only if IMU errors and estimated biases stay within their 3-sigma bounds. 3 sigma Level: must. docs.openvins.com 2026-11-09“Ensure that your estimated biases do not leave your 3-sigma bounds.”
  59. Compare calibrated rotation and translation with hand measurements. Level: should. docs.openvins.com 2026-11-09“Sanity check your final rotation and translation with hand-measured values.”
  60. Choose encoder resolution at least 2 to 4 times finer than the required position accuracy. 2 times Level: guide. support.maxongroup.com 2026-11-09
  61. For field-oriented control of a brushless motor, use at least a 500 count-per-turn incremental or an 11-bit absolute encoder, stiffly on the motor side. 500 cpt Level: guide. support.maxongroup.com 2026-11-09
  62. Assume backlash of 0.3 to 2 degrees for an ordinary gearhead until you have the gearbox figure. 0.3 ° Level: guide. support.maxongroup.com 2026-11-09
  63. To position a load through backlash, use a low-backlash drive or dual-loop control with a load-side encoder. Level: should. support.maxongroup.com 2026-11-09
  64. Compare the encoder step with the stated backlash before you quote accuracy. Example: 4096 pulses per turn beside 0.25 degrees of backlash. 0.25 ° Level: should. emanual.robotis.com 2026-11-09
  65. With no magnet displacement, no zero-programming and a 360° maximum angle, the AS5600 has ±1° system nonlinearity at 12-bit resolution. 1 degree Level: guide. files.seeedstudio.com 2026-11-09
  66. Convert absolute encoder bits to steps as 2 to the n per turn. Level: should. support.maxongroup.com 2026-11-09
  67. The Playground SO-101 driver uses an STS3215 with 4096 ticks per turn. Level: guide. raw.githubusercontent.com 2026-11-09“MOTOR_MODEL = "sts3215" TICKS_PER_TURN = 4096”
  68. 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
  69. On a controller's commutation encoder, fix the sense magnet to the rotor and not to a reducer or linkage output. Level: must. mjbots.github.io 2026-11-09“should be attached to the rotor (NOT the output of any reducer or linkage)”
  70. For the default onboard moteus encoder, centre it on the sense magnet with about a 2 mm air gap. 2 mm Level: must. mjbots.github.io 2026-11-09“has an air-gap of approximately 2mm”
  71. If using a 3D printed encoder mount, make sure it does not warp; PLA and other low temperature filaments warp under load at room temperature. Level: should. mjbots.github.io 2026-11-09“PLA and other low temperature filaments will warp under load at room temperature”
  72. For ams on-axis sensors, air gap changes flux density and linearity; x-y displacement also changes linearity. Level: should. dtsheet.com 2026-11-09“The magnetic flux density changes with bigger air-gaps. The linearity changes with mechanical displacements in x and y direction.”
  73. Check the stainless grade next to a sense magnet, because some grades are magnetic. Level: should. dtsheet.com 2026-11-09“stainless steel may also be used, but some grades are magnetic, they should be avoided”
  74. For an AS5048 with an 8 mm magnet displaced 500 µm in x and y, nonlinearity is ±0.7°. 0.7 Deg Level: guide. datasheet.chipsmall.com 2026-11-09
  75. Keep the field at the die within the sensor's stated range across the air gap tolerance. 30 mT Level: should. datasheet.chipsmall.com 2026-11-09
  76. Read the sensor's field diagnostics at run time: the flags indicate a weak or high magnetic field, and the maker recommends monitoring the magnitude value too. Level: should. datasheet.chipsmall.com 2026-11-09Review note: Source says the flags indicate a weak or high field and recommends monitoring magnitude; it never says they show a mechanical problem.
  77. Stay in the recommended field range on this sensor; signal-to-noise ratio is greater at greater field. 300 G Level: should. allegromicro.com 2026-11-09Review note: Source does not tie the 300 to 1000 G recommended range to best noise; it says signal-to-noise is greater at greater field, in a sentence cut off in the capture.
  78. For an AS5048 with an 8 mm magnet displaced 500 µm in x and y, nonlinearity is ±1.2° from -40 to 150°C. 1.2 Deg Level: guide. datasheet.chipsmall.com 2026-11-09
  79. If the shaft must be magnetic, put a non-magnetic spacer between it and the magnet. Level: should. dtsheet.com 2026-11-09“a possible solution is to place a nonmagnetic spacer between shaft and magnet”
  80. Reduce the effect of magnet misalignment with a smaller air gap and a larger magnet. Level: should. allegromicro.com 2026-11-09
  81. XM430-W350 Present Position resets to a value within one rotation at power-on, on reboot, and when Position Control Mode is set or torque turns on in it. Level: guide. emanual.robotis.com 2026-11-09
  82. A single-turn magnetic sensor keeps a turn count through power-off only if a low-power counter stays supplied. Level: should. allegromicro.com 2026-11-09Review note: Source says key-off sensors are 'often' powered from the battery; it does not state 'only if' or require anything for must.
  83. Budget dynamic angle error as speed times the sensor's propagation delay, and use a compensated output at speed. Level: should. datasheet.chipsmall.com 2026-11-09
  84. Plan one-time zero programming carefully; the burn command works at most 3 times. 3 times Level: should. files.seeedstudio.com 2026-11-09
  85. Run this off-axis ring encoder's self-calibration after installation to reach the stated accuracy. 0.02 ° Level: guide. rls.si 2026-11-09“System accuracy ±0.004° to ±0.020° / ±14 to ±72 arcsec after encoder self-calibration”Review note: Capture never mentions a hollow joint; it describes an off-axis encoder for limited installation space.
  86. IMU axes must form a right-handed frame; a driver that receives left-handed data flips the y axis before publishing. Level: must. ros.org 2026-11-09“all frames must be right handed. If any data is reported left handed the driver must convert it to right handed by inverting the y axis.”
  87. At rest with body z up, the accelerometer z output must read +g in m/s^2, because it reports specific force and not coordinate acceleration. Level: must. ros.org 2026-11-09“if the body z axis points upwards, its z axis should indicate +g. This data must be in m/s^2.”
  88. If a PX4 controller cannot sit at the centre of gravity, set EKF2_IMU_POS_X/Y/Z to its offset. Level: should. docs.px4.io 2026-11-09“If you can't mount the controller in this position, then you should configure the following parameters to set offset relative to the CoG”
  89. Calibrate misalignment across the full chain from IMU sensing axes to the frame you care about; the factory case alignment is not enough for high performance. Level: should. vectornav.com 2026-11-09
  90. Count the isolator as a filter in the control loop: a soft mount adds latency. Level: should. docs.px4.io 2026-11-09“A soft airframe or soft vibration mounting increases latency (they act as a filter).”
  91. Reduce the vibration source first; adjust software filters only as a last step. Level: should. docs.px4.io 2026-11-09“It is better to reduce the source of vibrations, rather than filtering them out in software.”
  92. Pick an accelerometer range with headroom for peak vibration; the BMI088 offers up to 24 g for this reason. 24 g Level: should. bosch-sensortec.com 2026-11-08
  93. Do not solder the exposed die pad of this IMU to the PCB; package stress changes its performance. Level: must. cdn.sparkfun.com 2026-11-08
  94. Calibrate the D435i IMU yourself: it ships without internal calibration, so idle gyro output is non-zero and accelerometer gravity is off. Level: must. raw.githubusercontent.com 2026-11-09“does not include internal calibration, which may manifest itself with non-zero angular velocities produced at idle mode”
  95. Compare IMUs by noise density, because a noise sigma depends on the output rate. Level: should. vectornav.com 2026-11-09
  96. Budget drift from gyro noise as random walk times the square root of time. Level: should. vectornav.com 2026-11-09
  97. Give in-run bias stability the most weight when selecting an IMU, as it sets the floor on bias knowledge. Level: should. vectornav.com 2026-11-09
  98. Turn-on bias varies between starts, so production calibration cannot remove it; an aided inertial navigation system can estimate it after each startup. Level: guide. vectornav.com 2026-11-09
  99. Budget attitude error from uncorrected gyro bias as bias times time. Level: should. cl.cam.ac.uk 2026-11-09Review note: Source gives only the linear error formula and calls bias compensation trivial; no requirement or stated failure, so must is over-strong.
  100. Budget temperature-driven bias separately; the bias stability figure excludes it. Level: should. cl.cam.ac.uk 2026-11-09
  101. Budget gyro offset change with temperature from the temperature coefficient. 0.015 °/s per K Level: should. bosch-sensortec.com 2026-11-08
  102. Do not feed static noise figures straight into a moving estimator: the model leaves out scale factor errors and temperature bias changes. Level: should. github.com 2026-11-09
  103. Check gyro acceleration sensitivity on a vibrating robot: oscillating acceleration can shift the gyro bias. Level: should. vectornav.com 2026-11-09
  104. Start camera-IMU calibration at about 20 Hz camera and 200 Hz IMU rate. 20 Hz Level: guide. github.com 2026-11-09
  105. Record at least 3 hours of still IMU data for an Allan analysis; longer gives better accuracy. 3 h Level: must. raw.githubusercontent.com 2026-11-09“You must record **at least** 3 hours of data.”
  106. Kalibr's authors recommend about 15 to 24 hours of stationary IMU data for Allan analysis. 15 hour Level: guide. github.com 2026-11-09
  107. OpenVINS recommends collecting upwards of 20 hours of stationary IMU data. 20 hours Level: guide. docs.openvins.com 2026-11-09“Collect a stationary dataset on the upwards of 20 hours”
  108. Do not rely on 2 to 3 percent distortion being harmless when you measure from the image; that figure holds only without measurement algorithms. 3 % Level: should. edmundoptics.com 2026-11-09“Typically, ±2 to 3% distortion is unnoticed in a vision system if measurement algorithms are not used.”
  109. Estimate rolling shutter skew from row read-out time, row count and image speed. Level: should. whitepapers.axis.com 2026-11-09Review note: Source only explains what sets the row time offset; no requirement or failure from not estimating skew, so must is over-strong.
  110. Do not expect shorter exposure to cure rolling shutter skew; it only reduces blur. Level: must. whitepapers.axis.com 2026-11-09
  111. Choose a global shutter where things move fast in the image, and pair it with short exposure. Level: should. whitepapers.axis.com 2026-11-09
  112. Do not use a short strobe with a rolling shutter unless the flash covers all rows, or you get a bright band. Level: must. whitepapers.axis.com 2026-11-09
  113. Keep exposure short enough that the target moves less than about one pixel during it. 1 pixel Level: should. assets-ctf.baslerweb.com 2026-11-09
  114. Do not expect global shutter to remove motion blur; only exposure does that. Level: must. assets-ctf.baslerweb.com 2026-11-09
  115. Model rolling-shutter effects in direct visual-inertial odometry; neglecting them degrades accuracy and robustness. Level: must. arxiv.org 2026-11-09
  116. Check depth imager shutter before you select a stereo module for a moving arm; modules in one family differ. Level: should. docs.clearpathrobotics.com 2026-11-08
  117. For RealSense D400 cameras, depth error grows with distance squared; work as close as MinZ allows. Level: should. dev.realsenseai.com 2026-11-09
  118. Use the ZED accuracy range, about 1 percent near and about 9 percent at maximum range, as the working budget. 9 % Level: should. stereolabs.com 2026-11-09
  119. Use the maker's per-range error: below 2 percent under 4 m, then below 4 percent to 7 m and below 6 percent to 10.5 m. 2 % Level: should. docs.luxonis.com 2026-11-09
  120. Read the D400 2 percent figure only with its scope: within 80 percent of the field and within the stated distance. 2 % Level: should. docs.clearpathrobotics.com 2026-11-08
  121. Trade baseline with care: a wider baseline improves depth accuracy but raises the minimum distance. Level: should. docs.luxonis.com 2026-11-09
  122. Calculate minimum stereo distance from focal length, baseline and maximum disparity. Level: should. docs.luxonis.com 2026-11-09Review note: Source only gives the formula as a worked OAK-D example; no requirement or failure, so must is over-strong.
  123. If D400 MinZ prevents getting closer, reduce depth X-resolution to reduce MinZ. Level: should. dev.realsenseai.com 2026-11-09
  124. Check exposure first when depth quality is poor, as both under and over exposure hurt. Level: must. dev.realsenseai.com 2026-11-09
  125. Set stereo depth gain low and adjust exposure first; gain adds noise that degrades depth. Level: should. dev.realsenseai.com 2026-11-09
  126. Provide texture for stereo: passive matching quality depends on visually distinct points. Level: should. raw.githubusercontent.com 2026-11-09“Any source of texture, natural or artificial will improve the accuracy significantly.”Review note: Source says texture improves accuracy significantly and a projector is useful; it does not require texture, so must is over-strong.
  127. Test white walls, plain surfaces and reflective areas, because they produce outliers in stereo depth. Level: should. stereolabs.com 2026-11-09Review note: Source says accuracy 'may also be reduced' on these surfaces and asks for no test, so must is over-strong.
  128. Design to the camera's ideal depth range, not its maximum. ZED Mini: 15 m maximum and 9 m ideal. 9 m Level: should. stereolabs.com 2026-11-09
  129. To reduce blur on ZED cameras, use manual mode and increase gain before exposure. Level: should. stereolabs.com 2026-11-09
  130. Treat depth value 0 as invalid data, not as a distance. Level: must. learn.microsoft.com 2026-11-09
  131. Saturated IR pixels lose phase information and are invalidated in Azure Kinect DK depth. Level: guide. learn.microsoft.com 2026-11-09
  132. Expect invalid depth at corners with a ToF camera, as light from more than one object reaches a pixel. Level: should. learn.microsoft.com 2026-11-09
  133. Read ToF operating range with its conditions: reflectivity, ambient IR, and error limits. 17 mm Level: should. learn.microsoft.com 2026-11-09
  134. Check ToF range for each mode. Narrow unbinned runs 0.5 to 3.86 m. 3.86 m Level: should. learn.microsoft.com 2026-11-09
  135. Use short high-quality USB3 cables for depth cameras, preferably under 1 m, and active repeaters beyond 2 m. 1 m Level: should. dev.realsenseai.com 2026-11-09
  136. Budget power for each depth camera; a camera with the projector on can reach or exceed 2 W. 2 W Level: should. dev.realsenseai.com 2026-11-09
  137. OpenCV hand-eye calibration needs at least 2 motions with non-parallel rotation axes, so at least 3 poses; many more are strongly recommended. 2 motions Level: must. docs.opencv.org 2026-11-09“A minimum of 2 motions with non parallel rotation axes are necessary to determine the hand-eye transformation.”
  138. Expect a first result after 5 samples; quality improves with more and typically plateaus at about 12 or 15. 12 samples Level: guide. moveit.github.io 2026-11-09“The calibration will improve significantly with a few more samples, and will typically plateau after about 12”Review note: Source says the plateau comes after about 12 or 15 samples; the record drops 15.
  139. Rotate between each pair of poses, and about more than one axis. 2 axes Level: must. moveit.github.io 2026-11-09“at least two rotation axes are needed to uniquely solve for the”
  140. Keep the target stationary in the robot base frame during eye-in-hand calibration. Level: must. moveit.github.io 2026-11-09“as the target is stationary in the robot's base frame, the hand-eye calibration can be estimated from a sequence of 5 or more poses”
  141. Maximise rotation between poses. Level: should. raw.githubusercontent.com 2026-11-09
  142. Keep the target close to the camera during hand-eye calibration. Level: should. raw.githubusercontent.com 2026-11-09
  143. Minimise translation between poses during hand-eye calibration. Level: should. raw.githubusercontent.com 2026-11-09
  144. Calibrate the robot if necessary or applicable. Level: should. raw.githubusercontent.com 2026-11-09Review note: Source is the bare tip 'Calibrate the robot if necessary / applicable'; the reason about kinematic error is not in it.
  145. Prefer a method that solves rotation and translation together when the data is noisy; the authors' comparison of three methods found it seems the most robust. Level: guide. arxiv.org 2026-11-09Review note: Source is not a survey: the authors compare their two methods with Tsai and Lenz, and say the method 'seems' most robust.
  146. Keep the target flat, or it cannot be localised reliably. Level: must. moveit.github.io 2026-11-09“The target must be flat to be reliably localized by the camera.”
  147. Protect small force-torque sensors during assembly; bolting them up can overload them. Level: must. ati-ia.com 2026-11-09
  148. Look up overload values per model and per axis. A Nano17 Titanium, for example, is rated to 160 N in x-y and 310 N in z. 160 N Level: must. ati-ia.com 2026-11-09
  149. Check combined loads, because a complex load can saturate the gauges while each axis is below its rating. Level: must. ati-ia.com 2026-11-09
  150. Take a bias or reference reading at the current temperature just before applying the load you want to measure. Level: should. ati-ia.com 2026-11-09
  151. Do not assume compensation removes zero drift: temperature compensation covers gain, and drift varies per unit. Level: should. ati-ia.com 2026-11-09
  152. Add temperature gain error to rated room-temperature accuracy for an overall estimate, if unloaded and loaded readings share a temperature; ATI non-Gamma transducers show 0.5% typical gain error at ±15°C from 22°C. 0.5 % Level: should. ati-ia.com 2026-11-09
  153. ATI non-Gamma transducers show 5% typical gain error at ±50°C from 22°C, bounded by their operational limits. 5 % Level: guide. ati-ia.com 2026-11-09
  154. Recalculate range, resolution and accuracy when you move the point of origin; published figures hold at the factory origin only. Level: must. ati-ia.com 2026-11-09
  155. Restrain the sensor cable close to its connector so repetitive motion does not reach the connector. Level: must. ati-ia.com 2026-11-09
  156. Rebias or compensate during long runs; the authors report that temperature drift can be considerable. Level: should. arxiv.org 2026-11-09
  157. Use a force-torque sensor's anti-alias filter, or provide your own where the system has none. Level: should. ati-ia.com 2026-11-09Review note: Source says the filter prevents aliasing and is absent on some models; 'provide your own' and must are not in it.
  158. Read a force-torque filter's group delay before closing a force loop on it. Level: should. ati-ia.com 2026-11-08
  159. Do not plan on lidar data inside the blind zone. Livox Mid-360: 0.1 m. 0.1 m Level: must. livoxtech.com 2026-11-08
  160. Treat returns from just outside the blind zone as reference only: precision is not guaranteed there. Level: must. livoxtech.com 2026-11-08
  161. Check minimum range for the exact model and firmware. Ouster OS1: 0.5 m in this datasheet. 0.5 m Level: should. cdn.shopify.com 2026-11-09
  162. Test any zero-blind-range claim on your own unit. 0 m Level: guide. hesaitech.com 2026-11-08
  163. At 100 klx, 1024×10 Hz and over 90% detection probability, OS1 rev7 ranges 90 m at 10% and 170 m at 80% Lambertian reflectivity. 90 m Level: guide. cdn.shopify.com 2026-11-09
  164. Read 2D scanner range by reflectance: this scanner drops from 10 m on white to 4 m at 10 percent reflectance. 4 m Level: should. hokuyo-aut.jp 2026-11-08Review note: Source only lists range for two reflectances; no requirement or failure, so must is over-strong.
  165. Check surrounding light limit, and avoid direct sun on this scanner. 80000 lx Level: must. hokuyo-aut.jp 2026-11-08
  166. Repeat temperature calibration when temperature moves more than 8 C from the last calibration. 8 degC Level: must. pololu.com 2026-11-09
  167. Size a ToF range sensor for bright light, not dark. VL53L1X long mode falls from 360 cm in the dark to 73 cm in strong ambient. 73 cm Level: should. pololu.com 2026-11-09Review note: Figures are right (360 dark, 73 strong ambient), but the source only tabulates them and requires no bright-light sizing, so must is over-strong.
  168. Choose short distance mode in bright light: range holds at about 135 cm. 135 cm Level: should. pololu.com 2026-11-09
  169. Repeat RefSPAD and crosstalk calibration whenever you add a cover glass over the sensor. Level: must. pololu.com 2026-11-09
  170. Load stored calibration data into the sensor at each start. Level: must. pololu.com 2026-11-09
  171. Select the timing budget for the distance mode; 20 ms works only in short mode, and 33 ms is the minimum for all modes. 33 ms Level: must. pololu.com 2026-11-09
  172. Check that the target fills the sensor's field of view, because the published range tables assume that. Level: should. pololu.com 2026-11-09
  173. Relate the test figures to your site: office light is near 5 kcps/SPAD and 200 kcps/SPAD is direct sun through a window. 5 kcps/SPAD Level: guide. pololu.com 2026-11-09
  174. Use the published field of view in your geometry check. 27 ° Level: should. pololu.com 2026-11-09
  175. Plan for offsets that change during a run: background calibration in fusion modes cannot be disabled. Level: should. bosch-sensortec.com 2026-11-08
  176. Test pitch and roll drift under vibration or constant motion, because the fusion ignores the accelerometer during motion. Level: must. bosch-sensortec.com 2026-11-08
  177. Test any sustained acceleration case, because the fusion can read it as gravity. Level: should. bosch-sensortec.com 2026-11-08Review note: The maker's own wording is 'it is recommended to test', so must is over-strong.
  178. Validate built-in fusion on robot motion; this algorithm was tuned for human motion. Level: should. bosch-sensortec.com 2026-11-08
  179. Use the game rotation vector where a heading jump would hurt and slow yaw drift is acceptable; it does not use the magnetometer. Level: should. ceva-ip.com 2026-11-08
  180. For simulated BNO08x motion with nominal calibration, gaming rotation-vector dynamic heading drift is 0.5°/min. 0.5 °/min Level: guide. ceva-ip.com 2026-11-08
  181. For simulated BNO08x motion with nominal calibration, rotation-vector error is 3.5° dynamic and 2.0° static. 3.5 ° Level: guide. ceva-ip.com 2026-11-08
  182. Avoid the accelerometer-and-magnetometer output on a fast or magnetically noisy robot: it responds slowly and errs in varying fields. Level: should. ceva-ip.com 2026-11-08
  183. Use a magnetometer fusion mode only where the surrounding field is stable; the maker warns that fields around electronics are usually much larger than the Earth's. Level: should. bosch-sensortec.com 2026-11-08Review note: Source says compass accuracy depends on field stability; it does not restrict use to stable fields, so 'only where' and must go too far.
  184. Mount the magnetometer away from power distribution and other magnetic sources. Level: should. ardupilot.org 2026-11-09“mounted on a mast up and away from magnetic source of interference including the power distribution board”
  185. Keep DC power leads short and twist them to limit field at the magnetometer. Level: should. ardupilot.org 2026-11-09“Keep the wires between the PDB, ESCs and battery as short as possible.”
  186. Do not treat offset-only calibration as complete; it corrects hard iron only. Level: should. docs.px4.io 2026-11-09“This type of calibration only estimates the offsets to compensate for a hard-iron effect.”
  187. Calibrate well away from strong magnetic objects, or their effect gets saved into the correction. Level: should. ceva-ip.com 2026-11-09
  188. Fuse a second orientation source only if its covariance is honest; if not, use only the more accurate source or fuse it differentially. Level: should. docs.ros.org 2026-11-09
  189. Transform IMU data to ENU before robot_localization; the node does not work with NED data. Level: must. docs.ros.org 2026-11-09
  190. For the lowest-cost IMUs, raise the noise parameters by 10 times or more in calibration. 10 times Level: guide. github.com 2026-11-09
  191. OpenVINS suggests inflating noise by about 10 to 20 times, and testing which factor works. 10 x Level: guide. docs.openvins.com 2026-11-09“Typically these noise values should be inflated by maybe 10-20x their values to take into account unmodelled errors”
  192. Note that rebiasing just before a load change is not practical on a floating-base robot, so plan an in situ method there. Level: guide. arxiv.org 2026-11-09
  193. Remember short distance mode limits range to about 1.3 m. 1.3 m Level: should. pololu.com 2026-11-09