Embedded Control and Motion Systems

Introduction

Embedded motion systems combine sensing, computation, actuation, and mechanical dynamics in a real-time feedback loop. Examples include robotic joints, automated guided vehicles, camera gimbals, CNC axes, drones, and prosthetic devices. Unlike a purely theoretical controller, an embedded controller must execute within a finite sampling period, use quantized and noisy measurements, respect actuator limits, and remain safe when communication or sensing fails.

This chapter develops the foundations needed to implement discrete-time feedback control, PID and state-space controllers, sensor fusion, trajectory generation, and hardware-in-the-loop (HIL) evaluation. The central design principle is that the algorithm, sampling process, sensors, processor, power electronics, and mechanical plant must be treated as one system.

Learning objectives

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

1. Explain sampling, discretization, quantization, computational delay, and numerical limitations in embedded control.
2. Implement and tune discrete-time PID controllers with anti-windup and output constraints.
3. Represent a plant in discrete state-space form and design a practical state-feedback or observer-based controller.
4. Combine measurements using complementary and Kalman-filter concepts.
5. Generate feasible position, velocity, and acceleration trajectories.
6. Evaluate stability, performance, robustness, and safety in real-time and HIL experiments.

1. The embedded motion-control loop

A typical digital motion system contains a reference generator, controller, power stage, actuator, mechanical plant, sensors, signal-conditioning circuits, and communication interfaces. At sampling instant k, the controller receives measurements, estimates the system state, calculates a command, and applies it until the next update.

Figure: Basic closed-loop architecture for an embedded motion system.

The sampling period Tₛ must be short enough to represent the relevant plant dynamics and to provide adequate phase margin, but a shorter period increases processor load, sensor traffic, and sensitivity to timing jitter. A common initial engineering choice is a control bandwidth at least ten times lower than the sampling frequency, although the final choice depends on delay, noise, actuator bandwidth, and model uncertainty.

Important embedded effects include:

• Quantization: finite ADC, encoder, and PWM resolution produce discrete measurement and command values.
• Latency: sensor conversion, communication, scheduling, and computation delay reduce phase margin.
• Jitter: variation in execution or sampling time introduces timing noise.
• Saturation: voltage, current, torque, speed, and position limits prevent an ideal command from being applied.
• Numerical range: fixed-point arithmetic can overflow; floating-point arithmetic can lose precision when very large and small values are combined.

A real-time task should have a measured worst-case execution time below its deadline, not merely a favorable average execution time. Priority inversion, interrupt interference, missed deadlines, and watchdog behavior must be considered during implementation.

2. From continuous models to discrete time

A continuous linear time-invariant model is commonly written as

Here x is the state vector, u is the input vector, y is the output vector, and A, B, C, and D are matrices of compatible dimensions. For zero-order-hold inputs, which remain constant between samples, the exact discrete model is

where

The matrix exponential is exact for the assumed linear model and zero-order hold. If A is nonsingular, an equivalent expression is B_d = A^-1(A_d - I)B, but the integral or numerical routines are preferable when A is singular or poorly conditioned.

For a scalar first-order plant

with a > 0 and constant input over a sample, the exact discrete coefficients are

Forward Euler is simpler:

It is computationally inexpensive but only an approximation. For the unforced scalar system ẋ = -a x, Euler gives x[k+1] = (1-aTₛ)x[k]. Stability requires |1 − aTₛ| < 1, hence 0 < aTₛ < 2. This condition illustrates why an apparently small sampling period is not optional: an unstable numerical model can result even when the physical continuous plant is stable.

The bilinear, or Tustin, transformation maps a continuous transfer function into discrete time using

It generally gives better frequency-domain behavior than forward Euler, but frequency warping can matter near the chosen bandwidth. Discretization should therefore be validated by comparing step response, frequency response, and closed-loop poles—not selected solely by convenience.

3. Discrete-time PID control

For a reference r[k] and measured output y[k], the error is e[k] = r[k] - y[k]. A practical discrete PID controller can be expressed as

with rectangular integration

Kₚ is dimensionless for a position-error loop when u and e have compatible normalized units, Kᵢ has units of inverse time, and K_d has units of time. In a physical system their units must be selected so every term has the units of actuator command.

The derivative term amplifies high-frequency measurement noise. A filtered derivative is therefore often used:

where 0 ≤ α < 1 is a smoothing coefficient. Derivative-on-measurement can avoid a sudden derivative kick when the reference changes, because the derivative responds primarily to measured motion rather than to a reference step.

Saturation and anti-windup

If the unconstrained command exceeds the actuator limit, the applied command is

Integral windup occurs when the integrator continues accumulating error while the actuator is saturated. A back-calculation anti-windup method modifies the integrator as follows:

T_aw is an anti-windup time constant. The correction drives the internal command toward the command that can actually be delivered. Conditional integration is another valid method: integration is disabled when saturation and error would drive the actuator further into saturation.

Worked example: digital PI control

Suppose a velocity loop uses Kₚ = 0.8, Kᵢ = 4 s⁻¹, Tₛ = 0.01 s, previous integral I[k−1] = 0.20, and current error e[k] = 0.50 in normalized units. Ignoring saturation and derivative action,

The proportional contribution is 0.8 × 0.50 = 0.40, so the unconstrained command is u[k] = 0.40 + 0.22 = 0.62. If the actuator limit is ±0.55, the applied command is 0.55. Without anti-windup, repeated positive error would keep increasing I even though the motor cannot produce the requested command.

PID tuning should be performed with safety limits active. A useful sequence is to identify approximate plant gain and time constant, select a sampling rate, tune proportional action conservatively, add integral action to remove steady-state error, and introduce derivative or feed-forward only if noise and bandwidth permit. The final tuning must be tested under load, reversals, disturbances, and sensor faults.

4. State-space control and observers

PID is intuitive and effective for many single-loop systems, but coupled multivariable mechanisms often benefit from state-space methods. A discrete model is

A state-feedback law is

where K is the feedback-gain matrix and N is a reference precompensator when needed. The closed-loop state transition is A_cl = A_d - B_dK. Stability of the linear discrete closed loop requires every eigenvalue of A_cl to lie strictly inside the unit circle: |λ_i(A_cl)| < 1. This is a local model-based condition; saturation, nonlinear friction, unmodeled flexibility, and sampling delay can still cause failure.

The pair (A_d, B_d) is controllable if its controllability matrix has full rank. For an n-state system,

must have rank n. Similarly, the pair (A_d, C) is observable if

has rank n. Controllability concerns whether inputs can move the states; observability concerns whether states can be reconstructed from outputs.

When not all states are measured, a discrete observer estimates them:

The estimation error evolves as e_x[k+1] = (A_d - LC)e_x[k]. Observer poles should be stable and are often faster than controller poles, but excessively fast observers amplify measurement noise and become sensitive to model error.

A Kalman filter provides a systematic observer when process and measurement disturbances are reasonably represented as zero-mean noise. The model is

where w and v have covariance matrices Q and R. The filter alternates prediction and correction. In simplified notation,

P is the estimation-error covariance. Q expresses uncertainty in the model and unmeasured disturbances; R expresses sensor noise. Incorrect covariance choices do not automatically make a filter unsafe, so innovation monitoring and fault handling remain necessary.

5. Sensor fusion for motion systems

No single sensor is ideal over all frequencies. An encoder may provide accurate low-frequency position but suffer quantization and differentiation noise. A gyroscope measures angular rate with good short-term response but bias drift. An accelerometer can provide a gravity-based inclination estimate at low frequency, but vibration and translational acceleration corrupt it.

A complementary filter exploits these characteristics. For an angle estimate θ, a simple form is

where ω is measured angular rate and 0 ≤ α ≤ 1. The integrated rate dominates rapid changes; the absolute sensor corrects long-term drift. The filter is easy to implement, but α is related to the desired crossover and should be validated experimentally.

Sensor fusion requires more than averaging. Measurements must be time-stamped, calibrated, transformed into a common coordinate frame, and checked for outliers. Bias estimation is often essential. A practical system should define behavior for stale data, impossible jumps, disconnected sensors, and inconsistent redundant measurements. Fused estimates should include validity status or uncertainty so higher-level control can reduce speed or enter a safe state when confidence is low.

6. Trajectory generation and motion profiles

A command step may be acceptable for a low-power laboratory motor but can cause excessive torque, vibration, or current in a mechanism. Trajectory generation creates a time history that respects position, velocity, acceleration, and sometimes jerk limits.

A trapezoidal velocity profile has acceleration, constant-velocity, and deceleration phases. For a move of distance Δq with symmetric acceleration magnitude a_max and velocity limit v_max, the triangular-profile peak velocity is

provided this value is no greater than v_max. The corresponding acceleration time is t_a = v_peak/a_max, and total move time is 2t_a. If √(a_maxΔq) > v_max, the profile is trapezoidal: t_a = v_max/a_max, cruise distance is Δq − v_max²/a_max, cruise time is that distance divided by v_max, and total time is 2t_a + t_c.

Worked trajectory example

Let Δq = 1.0 m, a_max = 2 m/s², and v_max = 1 m/s. The triangular candidate is

Since it exceeds the velocity limit, use a trapezoid. Each acceleration or deceleration phase covers

The cruise distance is 1.0 − 0.25 − 0.25 = 0.50 m. Thus t_a = 1/2 = 0.50 s, t_c = 0.50/1 = 0.50 s, and total duration is 1.50 s. The profile reaches the requested limits exactly and avoids a discontinuous velocity command.

For smoother mechanisms, a seventh-order or fifth-order polynomial can impose boundary conditions on position, velocity, and acceleration. A fifth-order polynomial is appropriate when position, velocity, and acceleration are specified at both endpoints. Its coefficients must be checked against actuator limits; satisfying boundary conditions alone does not guarantee feasible intermediate values.

In multi-axis motion, independent timing can produce path distortion. Coordinated trajectories use a common normalized path parameter so axes begin and end together. Collision checking and workspace limits should occur before the trajectory is sent to the low-level controller.

7. Stability, performance, and numerical limitations

For a discrete linear system, asymptotic stability requires closed-loop poles strictly inside the unit circle. A pole near 1 produces slow decay; a pole near −1 produces alternating responses. Frequency-domain margins, time-domain overshoot and settling time, actuator effort, and disturbance rejection should all be evaluated.

A useful second-order reference model has natural frequency ω_n and damping ratio ζ. Its continuous characteristic polynomial is

This model is a design reference, not a guarantee that a higher-order nonlinear mechanism will behave identically after discretization.

Numerical limitations include overflow, underflow, cancellation, finite word length, and accumulation of rounding error. Engineering practices include scaling variables to comparable magnitudes, using saturated arithmetic where appropriate, checking denominators before division, avoiding unnecessary subtraction of nearly equal quantities, and logging exceptional numerical states. Fixed-point implementations require explicit word-length and overflow analysis; floating point does not eliminate the need for range and precision analysis.

Derivative estimates should not be calculated from raw noisy position unless filtering and resolution are adequate. Encoder velocity may be estimated over a window, but the window adds delay. Every filter adds phase shift, and every delay must be included in stability analysis.

8. Hardware-in-the-loop evaluation

HIL testing connects the real controller to a real-time simulation of the plant. It allows faults, extreme loads, and repeatable scenarios to be tested before connecting the mechanism. A useful progression is:

1. Software-in-the-loop: controller code runs in a desktop simulation.
2. Processor-in-the-loop: compiled code runs on the target processor while the plant remains simulated.
3. HIL: target electronics exchange real-time signals with a plant simulator.
4. System test: the controller operates the physical mechanism under controlled conditions.

The HIL model must execute faster than real time with bounded latency. I/O scaling, sensor quantization, actuator saturation, communication delay, and fault injection should resemble the intended hardware. Test cases should include reference steps and ramps, reversals, load changes, sensor dropout, timing jitter, overcurrent, emergency stop, startup, shutdown, and recovery.

Record reference, measured output, estimated state, command before and after saturation, loop execution time, missed deadlines, and fault flags. Common performance measures include rise time, settling time, overshoot, steady-state error, RMS tracking error, peak current, and energy. For a sampled tracking error e[k], an RMS measure over N samples is

The units of e_RMS are the same as the measured error. HIL results do not prove physical safety by themselves: model error, wiring faults, thermal effects, structural resonance, and unexpected human interaction require physical validation.

9. Applied case: illustrative servo axis

Background

Consider an illustrative single-axis linear servo for a laboratory pick-and-place machine. The axis uses a motor-driven screw, an incremental encoder, a current-limited drive, and a microcontroller with a 1 kHz position-loop task. A faster inner current loop is assumed to be handled by the drive.

Problem

Direct position steps cause vibration and current saturation when the payload changes. Encoder differentiation is noisy, and a position-only PID controller accumulates integral error during acceleration. The engineering goal is accurate 0.40 m moves with limited velocity and acceleration, reliable behavior under a doubled payload, and safe response to encoder failure.

Analysis

The controller separates the problem into layers. A trajectory generator produces position and velocity references. A velocity or torque loop then tracks these references. The measured encoder position is filtered only enough to avoid excessive noise delay. The command is limited according to drive current and mechanical speed limits. The integral term is disabled or back-calculated during saturation.

A nominal discrete plant is identified from step tests and represented in state space using position and velocity as states. A state observer estimates velocity when direct differentiation is too noisy. The observer residual, defined as measured position minus predicted position, is monitored. A persistent residual beyond a calibrated threshold indicates a possible encoder or model fault.

Approach and solution

The design uses a jerk-limited or trapezoidal reference rather than a step, feed-forward velocity when available, and feedback gains selected from the nominal model. The sampling period and measured computation time are logged. HIL tests inject payload changes, encoder dropouts, delayed measurements, and current saturation. A watchdog disables the drive if the control task misses its deadline or if position exceeds a software envelope.

During commissioning, gains are increased gradually with a mechanical stop and emergency stop available. The axis is tested unloaded, at nominal load, and at the maximum permitted payload. The estimator is not allowed to conceal invalid sensor data: a failed encoder causes a controlled stop rather than continued closed-loop motion.

Outcome

The trajectory reduces peak acceleration demand and prevents the integral state from growing during current saturation. Observer-based velocity improves damping, while residual monitoring detects sensor disconnection. HIL testing reveals that an additional communication delay reduces phase margin; the final controller therefore uses a lower bandwidth and a stricter deadline. The case demonstrates that successful motion control is a systems-engineering result, not merely a choice of PID gains.

Learning points

• Separate trajectory generation from feedback correction.
• Treat saturation and delay as design conditions, not exceptional events.
• Use estimation to improve information, but retain independent fault checks.
• Validate timing, numerical behavior, and safety transitions as carefully as tracking accuracy.

10. Safety, ethics, and professional practice

Motion controllers can injure people and damage equipment. A safe design uses risk assessment, mechanical limits, current and thermal protection, emergency-stop architecture, controlled restart, and clearly defined safe states. A software stop should not be the only protection against a hazardous actuator; independent hardware protection may be required by the application and applicable machinery or functional-safety standards.

Engineers should document assumptions, calibration procedures, known failure modes, test evidence, and limits of validation. Data logs may contain operator or production information and should be protected appropriately. Machine-learning estimators can be useful for anomaly detection, but their uncertainty, training coverage, false-alarm behavior, and fallback mode must be assessed. An opaque model should not silently override a verified safety interlock.

Contemporary developments and best practices

Modern motion platforms increasingly combine model-based control, embedded optimization, time-sensitive networking, and edge analytics. These capabilities are valuable only when deterministic timing, cybersecurity, and verification remain adequate. Good practice includes version-controlled parameters, reproducible builds, unit tests for numerical routines, hardware fault injection, static analysis, calibration traceability, and acceptance tests linked to measurable requirements.

A practical implementation checklist is:

• Define units, coordinate frames, signs, limits, and sampling rates before coding.
• Select a discretization method consistent with the plant bandwidth and delay.
• Measure worst-case execution time and jitter.
• Implement saturation, anti-windup, initialization, and bumpless transfer.
• Validate sensor calibration, timestamps, transformations, and fault flags.
• Check closed-loop poles or margins using the actual discrete model.
• Test trajectories against velocity, acceleration, jerk, workspace, and collision limits.
• Use SIL, PIL, HIL, and physical tests as complementary evidence.
• Log enough information to reproduce failures without compromising privacy or safety.

Chapter Summary

Embedded motion control requires coordinated treatment of dynamics, sampling, numerical computation, sensing, actuation, and safety. Continuous models become discrete through zero-order-hold conversion or approximations such as Euler and Tustin; the selected sampling period and delay affect stability. Discrete PID control is practical when implemented with filtering, output limits, and anti-windup. State-space feedback and observers provide systematic methods for coupled systems and unmeasured states, while complementary and Kalman-filter ideas combine sensors with different noise and drift characteristics. Trajectory generation prevents infeasible commands by respecting velocity and acceleration limits. Stability must be assessed using the discrete closed-loop model, while HIL testing exposes timing, saturation, fault, and integration problems before full physical operation. Professional practice requires explicit assumptions, measurable requirements, independent safety functions, and documented verification.

Key Terms

Sampling period: Time between successive controller updates.

Zero-order hold: Assumption that a digital command remains constant between samples.

Quantization: Mapping of a continuous-valued signal to finite discrete levels.

Latency: Time delay between measurement, computation, and actuation.

Jitter: Variation in the timing of nominally periodic operations.

PID controller: Feedback controller using proportional, integral, and derivative actions.

Anti-windup: Method that prevents integral accumulation during actuator saturation.

Saturation: Limitation that constrains a command to an allowable range.

State-space model: Matrix representation using system states, inputs, and outputs.

Controllability: Ability to move system states through available inputs.

Observability: Ability to infer internal states from measured outputs.

Observer: Dynamic estimator of unmeasured system states.

Kalman filter: Model-based estimator using covariance-weighted prediction and measurement correction.

Sensor fusion: Combination of measurements to obtain a more useful estimate.

Trajectory generation: Construction of a time-varying motion command.

Trapezoidal profile: Motion profile with acceleration, constant velocity, and deceleration phases.

Hardware-in-the-loop: Real-time testing of a controller with a simulated plant.

Phase margin: Frequency-domain indication of tolerance to additional phase lag before instability.

Worst-case execution time: Maximum verified time required by a real-time task.

Review Questions

Short Answer Questions

1. Why does computational delay reduce the achievable control bandwidth?
2. State the discrete-time stability condition for the eigenvalues of a linear closed-loop system.
3. What is the purpose of anti-windup in a PID controller?
4. Distinguish controllability from observability.
5. Why are trajectory profiles preferable to abrupt position steps in many mechanisms?

Long Answer Questions

1. Derive the exact discrete model of a first-order plant under a zero-order-hold input and compare it with forward Euler discretization.
2. Explain the implementation of a filtered discrete PID controller, including saturation and back-calculation anti-windup.
3. Describe how a discrete observer estimates unmeasured states and discuss the trade-off in selecting observer speed.
4. Compare complementary filtering and Kalman filtering for combining gyroscope and accelerometer information.
5. Develop a verification plan covering software-in-the-loop, processor-in-the-loop, HIL, and physical testing for a robotic axis.

Analytical / Application Questions

1. A controller has Tₛ = 0.005 s, Kₚ = 1.2, Kᵢ = 6 s⁻¹, e[k] = 0.3, and I[k−1] = 0.10. Calculate the new integral state and unconstrained PI command, assuming no derivative action.
2. For a 2 m move with a_max = 4 m/s² and v_max = 2 m/s, determine whether the profile is triangular or trapezoidal and calculate its total duration.
3. A motion system becomes oscillatory after a networked sensor is introduced. Propose a structured diagnostic process involving timestamps, delay measurement, discrete-model analysis, HIL testing, and controller redesign.

Multiple Choice Questions with Answers

Question 1

For a zero-order-hold discrete model, what does the input do between two sampling instants?

A. It remains constant

B. It becomes zero

C. It changes linearly

D. It is replaced by white noise

Correct Answer: A. It remains constant

Question 2

For the scalar Euler model x[k+1] = (1 − aTₛ)x[k], which condition gives asymptotic stability for a > 0?

A. aTₛ = 0

B. aTₛ > 2

C. 0 < aTₛ < 2

D. aTₛ < −2

Correct Answer: C. 0 < aTₛ < 2

Question 3

What is the principal purpose of PID anti-windup?

A. To increase sensor resolution

B. To prevent integral accumulation during actuator saturation

C. To eliminate all measurement noise

D. To remove the need for output limits

Correct Answer: B. To prevent integral accumulation during actuator saturation

Question 4

A discrete linear closed loop is asymptotically stable when its poles are located where?

A. Outside the unit circle

B. On the imaginary axis

C. At z = 1 only

D. Strictly inside the unit circle

Correct Answer: D. Strictly inside the unit circle

Question 5

Which property describes whether inputs can move all states of a model?

A. Observability

B. Controllability

C. Quantization

D. Identifiability of sensors only

Correct Answer: B. Controllability

Question 6

In a complementary filter for angle estimation, what generally corrects long-term drift in integrated gyroscope data?

A. The absolute or low-frequency sensor measurement

B. A larger derivative gain only

C. PWM quantization

D. The actuator saturation limit

Correct Answer: A. The absolute or low-frequency sensor measurement

Question 7

For a move with Δq = 1 m, a_max = 2 m/s², and v_max = 1 m/s, what profile is required?

A. Triangular, because the peak velocity is 0.5 m/s

B. Triangular, because the peak velocity is 1.414 m/s

C. Trapezoidal, because the velocity limit is reached

D. Impossible, because acceleration is nonzero

Correct Answer: C. Trapezoidal, because the velocity limit is reached

Question 8

What is the main purpose of hardware-in-the-loop testing?

A. To replace all physical testing permanently

B. To test a real controller against a real-time simulated plant

C. To increase encoder resolution

D. To tune only mechanical bearings

Correct Answer: B. To test a real controller against a real-time simulated plant

Question 9

In a Kalman filter, what does the measurement covariance R primarily represent?

A. Uncertainty in the sensor measurement

B. The actuator's maximum voltage

C. The number of controller states

D. The sampling frequency

Correct Answer: A. Uncertainty in the sensor measurement

Question 10

If a PI controller has Kᵢ = 4 s⁻¹, Tₛ = 0.01 s, e[k] = 0.5, and I[k−1] = 0.20, what is I[k] without anti-windup correction?

A. 0.02

B. 0.20

C. 0.22

D. 0.40

Correct Answer: C. 0.22