Sensors, Instrumentation and Calibration

Introduction

Robots act in the physical world through a continuous loop of measurement, estimation, decision, and control. Actuators produce motion or force, but sensors provide the information needed to determine whether the intended result has been achieved. A robot without adequate sensing may move accurately in a nominal environment yet fail when friction, payload, lighting, wear, or object position changes.

This chapter develops the principles used to select, integrate, condition, calibrate, and fuse sensors in robotic systems. It considers position, velocity, force, torque, tactile, proximity, inertial, vision, and range sensors. Particular attention is given to uncertainty, sampling, bias, noise, linearity, repeatability, and practical calibration.

Learning Objectives

By the end of this chapter, the learner should be able to:

1. Explain the operating principles and limitations of major robotic sensors.
2. Select sensors according to range, resolution, bandwidth, accuracy, environment, and cost.
3. Design basic signal-conditioning and data-acquisition chains.
4. Perform calibration using reference measurements and estimate uncertainty.
5. Distinguish accuracy, precision, resolution, repeatability, and hysteresis.
6. Apply basic sensor-fusion methods under uncertainty.
7. Evaluate a practical robotic sensing architecture and its safety implications.

1. Measurement as a Robotic Function

A sensor converts a physical variable into a measurable signal. The complete instrumentation chain is usually:

physical quantity → sensing element → transducer → signal conditioning → analog-to-digital conversion → estimation and control

For example, a strain-gauge load cell detects deformation, a Wheatstone bridge converts deformation into a small voltage, an instrumentation amplifier increases the signal, an analogue-to-digital converter samples it, and a controller estimates contact force.

A useful sensor model is:

where x is the true physical quantity, y is the measured output, f(x) is the sensor response, b is systematic bias, and n is random noise. The model assumes that the measurement can be separated approximately into a deterministic response, a slowly varying error, and a random component. Calibration attempts to characterize or remove f and b, but cannot eliminate all uncertainty.

Important performance terms

Accuracy and precision are not interchangeable. A biased sensor can be precise but inaccurate; a noisy sensor can have good average accuracy but poor precision.

2. Position and Displacement Sensors

2.1 Encoders

Rotary encoders measure angular position. Incremental encoders produce pulses as a shaft rotates. Quadrature channels A and B permit direction detection because their phase differs by approximately 90°. A separate index pulse can identify one reference position. Incremental encoders lose absolute position after power loss unless the system performs homing.

Absolute encoders provide a coded output representing shaft angle. Single-turn devices measure one revolution, while multi-turn devices also record revolutions. They retain position information but generally cost more and require careful communication and error checking.

If an encoder has N counts per revolution and quadrature decoding provides four count events per cycle, the angular increment per count is:

Here ∆θ is in radians. This is an ideal quantization limit; mechanical backlash, eccentricity, interpolation error, and installation errors may dominate in practice.

For a linear joint driven by a leadscrew with pitch p metres per revolution:

This assumes direct transmission, negligible compliance, and correct quadrature counting.

2.2 Resolvers, potentiometers, and linear transducers

A resolver uses sinusoidal and cosinusoidal signals to determine shaft angle and is robust in harsh industrial environments. Potentiometers are inexpensive and simple, but their contact wear and limited life make them less suitable for high-cycle precision applications. Linear variable differential transformers (LVDTs) provide contactless displacement measurement with good repeatability, although they require excitation and demodulation electronics.

2.3 Selection considerations

Position-sensor selection must consider joint type, required accuracy, travel, speed, environmental contamination, cable routing, backlash, and whether absolute recovery after power interruption is required. The sensor should measure the variable that matters at the controlled location. A motor encoder may not reveal output-shaft error caused by gearbox compliance.

3. Velocity and Acceleration Measurement

Velocity may be measured directly by a tachometer or estimated from position. A simple finite-difference estimate is:

where T_s is the sampling period. Differentiation amplifies high-frequency noise, so practical systems use filtering or observer-based estimation. A moving-average filter with window length M is:

This reduces random noise when samples are independent, but introduces delay and attenuates rapid changes.

Accelerometers measure specific force rather than simply “position acceleration.” Their output includes gravity, vibration, bias, and noise. Gyroscopes measure angular rate and are commonly integrated to estimate orientation. Integration causes bias accumulation: a constant rate bias b produces an angular error approximately equal to bt after time t. Consequently, inertial sensors are excellent for short-term motion estimation but require correction from other sensors for long-term stability.

4. Force and Torque Sensing

Force sensors commonly use strain gauges, piezoelectric elements, capacitive structures, or optical methods. A strain gauge changes resistance when stretched. In a Wheatstone bridge, small resistance changes generate a differential voltage that is amplified by an instrumentation amplifier.

A six-axis force–torque sensor measures three force components and three moment components:

The sensor output vector r is related to the wrench w through a calibration matrix C:

Here r₀ is the zero-load output. The matrix accounts for sensitivity, scale factors, and cross-axis coupling. Its use assumes the sensor remains within its calibrated range and that temperature and mounting conditions are sufficiently similar to calibration conditions.

Force sensing supports collision detection, compliant assembly, polishing, grasp control, and human–robot interaction. Piezoelectric sensors offer high bandwidth but are poorly suited to measuring static force because charge leakage causes drift. Strain-gauge sensors measure static and dynamic loads but require temperature compensation and careful mechanical mounting.

5. Tactile, Proximity, and Contact Sensors

Tactile sensors detect local contact, pressure distribution, slip, or surface texture. Resistive, capacitive, piezoresistive, optical, and magnetic technologies are used in robotic fingertips, grippers, and artificial skin. A tactile array can identify contact location and approximate pressure distribution, but spatial resolution, hysteresis, durability, and wiring complexity are important limitations.

Proximity sensors detect objects without physical contact. Inductive sensors detect conductive metals; capacitive sensors detect changes in electric field and can detect many non-metallic materials; photoelectric sensors use emitted and reflected or interrupted light; ultrasonic sensors infer distance from acoustic time of flight.

A time-of-flight range estimate is:

where d is target distance, c is propagation speed, and t is round-trip travel time. The factor 2 applies to an emitted pulse that travels to the target and returns. The equation assumes the propagation speed is known and that the detected echo corresponds to the intended target.

6. Vision and Range Sensing

Cameras provide rich information such as shape, colour, texture, and object identity. A pinhole camera model is:

where (X,Y,Z) is a point in camera coordinates, (u,v) is its image position, f is focal length in pixel units, and (c_x,c_y) is the principal point. This ideal model ignores lens distortion; real systems require intrinsic calibration for focal length, principal point, and distortion coefficients.

Stereo vision estimates depth from disparity. If two cameras have baseline B and focal length f, depth is approximately:

where d_s is image disparity. Depth uncertainty increases rapidly as disparity becomes small, so distant or textureless objects are difficult to measure accurately.

Structured-light and time-of-flight cameras provide depth more directly but can be affected by sunlight, reflective surfaces, multipath effects, transparent materials, and interference between sensors. Vision systems also require a calibrated transformation between camera coordinates and robot coordinates. This is known as hand–eye or extrinsic calibration.

7. Signal Conditioning and Data Acquisition

Sensor outputs may be weak, noisy, nonlinear, or incompatible with controller input. Signal conditioning can include amplification, bridge completion, isolation, excitation, linearization, filtering, protection, and temperature compensation.

Sampling must be fast enough to represent the information of interest. The Nyquist condition requires:

where f_s is sampling frequency and f_max is the highest significant signal frequency. In practice, sampling substantially above this minimum is preferred, and an analogue anti-aliasing low-pass filter is applied before conversion. Aliasing occurs when high-frequency content appears incorrectly as lower-frequency content after sampling.

The quantization step of an ideal N-bit ADC with input span V_range is approximately:

The quantization error is commonly bounded by ±q/2 under the ideal rounding model. Real ADC performance is also limited by nonlinearity, reference instability, electrical noise, and effective number of bits.

Grounding, shielding, twisted-pair wiring, differential measurement, isolation, and physical separation from motor cables are often as important as the nominal sensor specification. A filter should not be selected solely for noise reduction: excessive delay can destabilize a feedback loop or reduce collision-response performance.

8. Calibration and Error Analysis

Calibration establishes the relationship between sensor output and a reference quantity. A basic linear model is:

where a is sensitivity and b is offset. Given two reference points (x₁,y₁) and (x₂,y₂),

The corrected estimate is x_hat = (y − b)/a, provided a ≠ 0. This two-point method is suitable only when the response is sufficiently linear and stable.

For multiple measurements, least-squares linear regression estimates a and b by minimizing the sum of squared residuals:

Residual analysis should follow calibration. A curved residual pattern indicates nonlinearity; increasing spread may indicate heteroscedastic noise; separate increasing and decreasing curves indicate hysteresis.

Calibration should include:

1. A traceable reference or reference instrument.
2. Zeroing under the correct unloaded or neutral condition.
3. Multiple points across the operating range.
4. Increasing and decreasing tests when hysteresis is possible.
5. Temperature, mounting, and cable conditions representative of use.
6. Independent validation data not used to fit the calibration.
7. A documented calibration interval and acceptance criterion.

Uncertainty should be reported rather than hidden. If independent quantities have standard uncertainties u₁, u₂, …, and the output is g(z₁,z₂,…) then a first-order combined standard uncertainty is:

This assumes small uncertainties, a locally linear model, and negligible covariance. If quantities are correlated, covariance terms must be included. An expanded uncertainty is often reported as U = ku_c, where k is a coverage factor selected according to the reporting convention and confidence requirement.

9. Sensor Fusion Under Uncertainty

No sensor is ideal. Encoders provide excellent relative position but may suffer from backlash; inertial sensors provide rapid motion information but drift; cameras provide global geometric information but may fail under poor lighting; force sensors detect contact but can be disturbed by vibration.

Sensor fusion combines complementary measurements. A simple weighted estimate for two independent measurements z₁ and z₂, with variances σ₁² and σ₂², is:

The lower-variance measurement receives greater weight. Independence and correctly characterized variances are required for this expression.

The Kalman filter generalizes this principle for dynamic systems. A state model is written as:

where x_k is the state, u_k the input, z_k the measurement, A, B, and H system matrices, and w_k and v_k process and measurement noise. The classical Kalman filter assumes a linear model and zero-mean Gaussian noise with known covariance. Extended and unscented variants address nonlinear models, but they do not eliminate poor calibration or incorrect noise assumptions.

A practical fusion architecture may combine wheel or joint encoders with an IMU for short-term motion, and vision or external tracking for drift correction. Time synchronization and coordinate-frame calibration are essential. A mathematically sophisticated filter cannot compensate for a timestamp error or a camera mounted with an incorrectly measured transform.

10. Worked Example: Encoder Resolution and Calibration

An incremental encoder produces 2,048 cycles per revolution. Quadrature decoding counts four transitions per cycle. A gearbox has a 100:1 reduction, and the encoder is mounted on the motor shaft. The ideal output-shaft angular increment is:

Thus:

This is the quantization increment, not guaranteed absolute accuracy. Gearbox backlash, shaft compliance, encoder eccentricity, and transmission error may be much larger.

Suppose a force sensor produces 0.42 V at 0 N and 2.42 V at 100 N. The linear sensitivity is:

The offset is b = 0.42 V. If the measured output is 1.22 V, the estimated force is:

The result is valid only within the calibrated range and assuming negligible hysteresis and temperature drift.

11. Applied Case: Vision-Guided Robotic Pick-and-Place

Background. An industrial robot must pick randomly oriented components from a conveyor. A camera is mounted above the work area, while joint encoders report robot configuration.

Problem. Initial trials show missed picks. The camera detects parts accurately in image coordinates, but the robot approaches positions displaced by 8–15 mm. Errors vary across the work envelope.

Analysis. The image-to-robot transformation was estimated from only one reference point. This cannot reliably determine translation, rotation, and scale. Lens distortion was also ignored, and camera timestamps were not synchronized with conveyor motion.

Approach. The engineering team performs intrinsic camera calibration using a calibration target, then estimates the planar camera-to-conveyor transform from multiple surveyed points. Robot hand–eye geometry is verified using several poses. Conveyor speed is measured, and the target position is extrapolated to the expected pick time. Encoder and camera timestamps are synchronized. A validation set of parts at locations different from the calibration points is used to assess residual error.

Outcome. Position error is reduced to within the gripper’s allowable capture region. A proximity sensor confirms part presence, and tactile or motor-current feedback detects failed grasping. The system is designed to stop or request operator intervention when image confidence, calibration validity, or sensor health falls below specified limits.

Learning points. Calibration must represent the full operating geometry; coordinate frames and timing are as important as sensor resolution; independent validation is essential; and redundant sensing improves fault detection rather than merely increasing data volume.

12. Limitations, Safety, and Professional Practice

Sensor failure can cause unsafe motion. Failure modes include frozen output, intermittent wiring, saturation, drift, implausible jumps, and common-cause failure affecting redundant sensors. Safety-related systems should use diagnostics, plausibility checks, watchdogs, safe limits, and an appropriate risk assessment. A sensor used for personnel protection must meet applicable machinery-safety requirements; a general-purpose camera or encoder should not automatically be treated as a safety-rated device.

Data governance also matters. Vision systems may capture workers or sensitive environments. Access control, retention limits, secure communication, and transparent operational policies should be considered. Machine-learning perception systems require representative validation, monitoring for distribution shift, and a safe response when confidence is low. AI-based detection should complement—not silently replace—well-defined safety functions.

Contemporary developments include event-based cameras for high-speed motion, flexible tactile skins, fibre-optic force sensing, solid-state lidar, magnetic encoders with improved immunity to contamination, and self-diagnosing sensor networks. These technologies expand capability but introduce new calibration, cybersecurity, interoperability, and maintenance requirements.

Best-Practice Checklist

• Define the physical variable and required control or safety decision before selecting a sensor.
• Compare range, resolution, accuracy, bandwidth, latency, drift, environmental rating, and lifetime.
• Locate the sensor as close as practical to the variable being controlled.
• Condition and sample signals without introducing unacceptable delay or aliasing.
• Calibrate over the actual operating range and environmental conditions.
• Separate calibration data from validation data.
• Record uncertainty, calibration date, reference standard, and acceptance criteria.
• Synchronize clocks and document coordinate frames.
• Fuse sensors according to their uncertainty and failure characteristics.
• Design explicit fault detection and safe fallback behavior.

Chapter Summary

Robotic sensing transforms physical variables into information used for control, estimation, and safety. Position and velocity sensors support motion control; force, torque, and tactile sensors enable contact interaction; proximity and range sensors support detection and distance measurement; inertial sensors provide fast motion information; and vision provides rich geometric and semantic data. Every sensor has limitations involving range, noise, drift, latency, bandwidth, environmental sensitivity, and calibration.

Signal conditioning, appropriate sampling, anti-aliasing, and careful grounding are necessary for trustworthy data. Calibration establishes sensor response and exposes bias, nonlinearity, hysteresis, and drift. Uncertainty analysis distinguishes what is known from what is estimated. Sensor fusion combines complementary measurements, but depends on valid models, covariance assumptions, coordinate transformations, and time synchronization. In professional robotics, sensing is therefore not merely a component-selection task; it is an integrated engineering discipline involving measurement science, control, software, safety, and maintenance.

Key Terms

Sensor: Device that measures a physical quantity.

Transducer: Device that converts one form of energy or physical quantity into another measurable form.

Accuracy: Closeness of a measurement to the true value.

Precision: Closeness among repeated measurements.

Resolution: Smallest detectable change.

Sensitivity: Output change per unit input change.

Bias: Systematic measurement offset.

Drift: Time- or environment-dependent change in output.

Hysteresis: Dependence of output on measurement history.

Encoder: Device for measuring position or motion, commonly rotational.

Inertial measurement unit: Combination of accelerometers and gyroscopes, often with a magnetometer.

Wrench: Combined force and torque vector.

Tactile sensor: Sensor that measures contact, pressure, or related surface interaction.

Proximity sensor: Sensor that detects nearby objects without direct contact.

Aliasing: Misrepresentation of high-frequency content caused by insufficient sampling.

Calibration: Establishment of a measurement relationship against a reference.

Uncertainty: Quantified doubt associated with a measurement result.

Sensor fusion: Combination of multiple sensor measurements to estimate a state.

Hand–eye calibration: Estimation of the geometric relationship between a robot and a camera.

Review Questions

Short Answer Questions

1. Distinguish accuracy, precision, and resolution.
2. Why does differentiating position amplify noise?
3. State two advantages and two limitations of inertial sensors.
4. What is the purpose of an anti-aliasing filter?
5. Why is independent validation data needed after calibration?

Long Answer Questions

1. Compare incremental and absolute encoders for an industrial robot joint.
2. Explain the calibration and uncertainty-analysis process for a six-axis force–torque sensor.
3. Describe the operating principles and limitations of ultrasonic, optical, and inductive proximity sensors.
4. Explain how camera intrinsic and extrinsic calibration support robotic manipulation.
5. Discuss how signal conditioning, sampling, grounding, and filtering affect closed-loop robot performance.

Analytical / Application Questions

1. A robot uses an encoder, IMU, and camera for localization. Propose a fusion strategy and identify the dominant failure modes of each sensor.
2. Design a calibration experiment for a force sensor that must operate from 0 to 200 N across a changing-temperature factory environment.
3. A vision-guided pick system has accurate image detections but inconsistent robot pick locations. Develop a diagnostic sequence covering geometry, timing, calibration, and mechanical causes.

Multiple Choice Questions with Answers

Question 1

An incremental encoder has 1,000 cycles per revolution and is decoded using four transitions per cycle. What is its ideal angular increment per count?

A. 0.00025π rad

B. 0.0005π rad

C. 0.001π rad

D. 0.004π rad

Correct Answer: A. 0.00025π rad

Question 2

Which sensor is generally best suited to measuring static force over long periods?

A. Piezoelectric force sensor

B. Strain-gauge load cell

C. Ultrasonic proximity sensor

D. Gyroscope

Correct Answer: B. Strain-gauge load cell

Question 3

A time-of-flight distance sensor measures round-trip travel time. Why is the propagation distance divided by two?

A. The signal is amplified twice

B. The target reflects half the energy

C. The signal travels to the target and back

D. The sensor samples twice per cycle

Correct Answer: C. The signal travels to the target and back

Question 4

What is the principal danger of sampling a signal below the required rate without suitable filtering?

A. Hysteresis

B. Aliasing

C. Backlash

D. Quantization removal

Correct Answer: B. Aliasing

Question 5

A sensor output changes from 0.5 V at 0 N to 2.5 V at 100 N. What is its linear sensitivity?

A. 0.005 V/N

B. 0.02 V/N

C. 2 V/N

D. 50 V/N

Correct Answer: B. 0.02 V/N

Question 6

Which statement about inertial sensors is correct?

A. They never experience drift

B. Gyroscope bias can accumulate into orientation error

C. Accelerometers directly measure absolute position

D. They do not require coordinate-frame definitions

Correct Answer: B. Gyroscope bias can accumulate into orientation error

Question 7

In a weighted estimate using independent measurements, which measurement receives greater weight?

A. The one with larger variance

B. The one with smaller variance

C. The older measurement regardless of variance

D. The measurement with the largest numerical value

Correct Answer: B. The one with smaller variance

Question 8

What does hand–eye calibration determine?

A. The ADC quantization step

B. The relationship between a robot and a camera coordinate frame

C. The operating temperature of a motor

D. The encoder’s pulse frequency only

Correct Answer: B. The relationship between a robot and a camera coordinate frame

Question 9

Which residual pattern most strongly suggests sensor nonlinearity after fitting a straight calibration model?

A. Random residuals centred around zero

B. A curved systematic residual pattern

C. A constant zero residual

D. A timestamp that is exactly synchronized

Correct Answer: B. A curved systematic residual pattern

Question 10

A 12-bit ADC has an input span of 4.096 V. What is its ideal quantization step?

A. 0.001 V

B. 0.004 V

C. 0.012 V

D. 4.096 V

Correct Answer: A. 0.001 V