Classical Controller Design

Introduction

Classical controller design provides the practical foundation for regulating mechanical systems. A controller compares a desired reference with a measured output and generates an actuator command that reduces the error. In robotics, examples include controlling joint angle, motor speed, end-effector position, force, and temperature in actuators or manufacturing equipment.

Classical methods are based largely on transfer functions, frequency response, time-domain specifications, and experimentally tunable controller structures. They remain important even when modern state-space, adaptive, or learning-based methods are available because they are understandable, computationally efficient, and widely supported by industrial hardware.

A useful feedback loop contains a reference input r(t), an error e(t), a controller C(s), a plant G(s), a sensor, and an actuator. For unity negative feedback, the closed-loop transfer function is

where s is the Laplace variable, R(s) is the reference, and Y(s) is the output. This expression assumes linear, time-invariant components and ideal unity feedback. Real systems also contain disturbances, sensor noise, actuator limits, delays, and parameter uncertainty.

Learning Objectives

After studying this chapter, the reader should be able to:

• explain the operation and limitations of on–off, proportional, integral, derivative, PID, lead, and lag controllers;
• select controller structures for common robotic and mechanical applications;
• tune PID and lead–lag controllers using time-domain and frequency-domain reasoning;
• analyze steady-state error, stability, bandwidth, overshoot, and robustness;
• explain actuator saturation, integral windup, derivative noise, and anti-windup methods;
• design a practical cascade control system and evaluate engineering trade-offs.

1. Feedback and Performance Requirements

The error signal is

where y_m(t) is the measured output. With a sensor transfer function H(s), the general closed-loop relationship is

The characteristic equation is

and its roots determine closed-loop stability for a linear model. A stable continuous-time system has poles with negative real parts. Stability alone is insufficient: a system may be stable but too slow, excessively oscillatory, inaccurate, noisy, or demanding more actuator effort than the hardware can provide.

Common performance specifications include rise time, settling time, overshoot, steady-state error, bandwidth, disturbance rejection, and control effort. For a standard underdamped second-order model,

where ω_n is the natural frequency in rad/s and ζ is the damping ratio. For 0 < ζ < 1, the approximate percent overshoot for a unit-step response is

This relation applies only to a dominant second-order response with negligible additional poles and zeros. Increasing ζ generally reduces overshoot, while increasing ω_n generally makes the response faster but may increase noise sensitivity and actuator demand.

For a 2% settling-time criterion, a commonly used approximation is

again assuming dominant second-order behavior.

2. On–Off and Proportional Control

On–off control

An on–off controller switches the actuator between two states:

where u(t) is the command and U_max and U_min are actuator limits. A dead band is often introduced so that switching occurs only when the error exceeds a specified range. Without hysteresis, measurement noise can cause rapid switching, called chattering. On–off control is inexpensive and suitable for thermostats, limit-position mechanisms, and simple safety functions, but it cannot normally hold a continuously varying output precisely.

Proportional control

The proportional law is

where K_p is the proportional gain. In the Laplace domain, C(s) = K_p. Increasing K_p usually reduces error and increases speed, but excessive gain can produce oscillation or instability.

For a type-0 unity-feedback system with plant static gain K and a step reference of amplitude A, the steady-state error under proportional control is

provided the closed loop is stable. Thus proportional control reduces, but generally does not eliminate, step error for a type-0 plant. A high gain may be impractical because of actuator saturation, structural flexibility, friction, and unmodeled dynamics.

3. Integral, Derivative, and PID Control

Integral action

Integral control accumulates error:

where K_i has units appropriate to the plant and controller input. In transfer-function form,

Integral action increases system type by one and can eliminate steady-state error for constant references or disturbances, assuming stability and sufficient actuator authority. It also adds phase lag and can increase overshoot and settling time.

Derivative action

Derivative action responds to the rate of error change:

with transfer function C_D(s) = K_d s in an ideal model. Derivative action adds predictive damping, but ideal differentiation amplifies high-frequency measurement noise and is not physically realizable without filtering.

A practical derivative term is

where T_f is the derivative-filter time constant. The filter makes the high-frequency gain finite. Derivative action is often applied to the measured output rather than the error to avoid a sudden derivative kick when the reference changes.

PID controller

The ideal parallel PID controller is

or

A practical filtered form is

An alternative parameterization is

where T_i is integral time, T_d is derivative time, and N is a filter factor, commonly between about 5 and 20 in practical applications. The parameter relations are K_i = K_p/T_i and K_d = K_pT_d.

The three actions have distinct roles:

4. PID Tuning

Tuning should begin with a safe plant model or a low-gain experiment. A useful sequence is:

1. establish actuator limits, sensor range, and safe operating conditions;
2. tune proportional action for an acceptable speed and damping;
3. add integral action gradually to remove residual error;
4. add filtered derivative action only if additional damping or anticipation is needed;
5. test set-point changes, disturbances, saturation, noise, and parameter variation.

Model-based tuning

Suppose a motor-position plant is approximated by a first-order-plus-dead-time model,

where K is process gain, T is the time constant, and L is dead time. The model is useful for initial tuning but does not capture all flexible modes or nonlinear friction. Several empirical tuning rules exist, including rules based on desired closed-loop time constant. Their numerical parameters depend on the selected rule and robustness objective; therefore, they should be treated as starting points rather than universal answers.

Frequency-response tuning

The loop transfer function is

The gain crossover frequency is the frequency at which |L(jω)| = 1. The phase margin is

where ω_gc is the gain crossover frequency. The gain margin is the reciprocal of the loop magnitude at the phase-crossover frequency, expressed in decibels as

These definitions apply to a negative-feedback loop under the usual Nyquist interpretation. Larger margins generally indicate greater robustness, although the appropriate values depend on delay, uncertainty, resonance, and performance requirements. A design with high bandwidth but small margins may be fast yet fragile.

5. Lead, Lag, and Lead–Lag Compensation

Lead compensation

A lead compensator may be written as

Its zero is at −1/T and its pole is at −1/(αT), which lies farther to the left because α < 1. The compensator contributes positive phase over a frequency range and can increase phase margin and bandwidth.

The maximum phase lead is

and occurs at

where ω_m is in rad/s when T is in seconds. These equations assume the standard lead form and positive α and T. Lead compensation is useful when a robotic joint requires faster tracking but has insufficient phase margin.

Lag compensation

A lag compensator is commonly written as

The pole is closer to the origin than the zero. Lag compensation increases low-frequency loop gain relative to high-frequency gain, improving steady-state accuracy while usually reducing bandwidth and adding phase lag. It is useful when accuracy is more important than speed and when measurement noise or actuator limitations make a high proportional gain undesirable.

Lead–lag compensation

A lead–lag controller combines both effects:

with 0 < α < 1 for the lead section and β > 1 for the lag section. Design should separate their frequency ranges where possible. The lead section is normally placed near the desired crossover frequency; the lag section is placed at lower frequencies so that its phase penalty near crossover is limited.

6. Cascade Control

Cascade control uses an outer loop around one or more faster inner loops. A typical robotic actuator may use an outer position loop, a middle velocity loop, and an inner current or torque loop. The inner loop must be substantially faster than the outer loop; otherwise, the outer controller acts on dynamics that have not settled.

For a two-loop system, the inner controller C₂(s) regulates an inner plant G₂(s). The resulting equivalent inner closed-loop transfer function becomes part of the outer plant. A practical design sequence is:

1. close and tune the innermost loop;
2. verify its stability, bandwidth, and current or torque limits;
3. close the next outer loop and tune it using the measured inner-loop behavior;
4. repeat outward;
5. test interactions, saturation, sensor delays, and disturbances.

Cascade control improves disturbance rejection when the disturbance enters inside the inner loop. For example, motor torque disturbances can be rejected by a current loop before they significantly affect position. However, cascade control requires additional sensors, calibration, bandwidth separation, and careful management of limits.

7. Saturation, Windup, Noise, and Practical Nonlinearities

Actuator saturation

An actuator cannot produce unlimited command. A saturated command is

where u_c is the controller's unconstrained output. Saturation makes the system nonlinear. A controller designed from a linear model may therefore perform very differently during large commands or disturbances.

Integral windup

If the actuator saturates while error persists, the integral term continues to accumulate. When the error changes sign, the stored integral must first unwind, producing excessive overshoot and slow recovery.

A practical back-calculation anti-windup law is

where x_I is the integrator state, K_aw is the anti-windup gain, u_c is the unsaturated command, and u_sat is the limited command. The correction drives the integrator toward a value consistent with the actuator output. Another method is conditional integration: stop integrating when the actuator is saturated and the error would drive it further into saturation.

Anti-windup should account for rate limits as well as amplitude limits. A motor may have sufficient voltage range but insufficient acceleration or current slew rate.

Measurement noise

Derivative action is especially sensitive to noise. Good practice includes sensor filtering, derivative filtering, appropriate sampling, shielding, grounding, and avoiding unnecessarily high bandwidth. Filters introduce phase delay, so their effect must be included in stability analysis. Excessive filtering can make the controller slow and can reduce disturbance rejection.

Uncertainty and unmodeled dynamics

Mechanical systems contain friction, backlash, compliance, resonance, load variation, gearbox elasticity, and delays. A controller should not be tuned solely for a nominal model. Robustness should be checked over plausible mass, inertia, friction, payload, and sensor-delay ranges. Gain scheduling may be useful when operating conditions vary substantially, but transitions between schedules must be smooth and safe.

8. Worked Example: PI Control of a First-Order Plant

Consider the normalized plant

with unity feedback and a PI controller

The closed-loop characteristic equation is obtained from 1 + C(s)G(s) = 0:

Therefore,

For K_p = 3 and K_i = 2,

The poles are

so the poles are approximately −0.586 and −3.414 rad/s. Both poles are negative, hence the linear closed loop is stable. Because integral action is present and the loop is stable, the steady-state error to a constant step reference is zero. The slow pole near −0.586 dominates the response, so the approximate 2% settling time is about 4/0.586 ≈ 6.8 s. This estimate is only approximate because the system is not a standard second-order system with a repeated or clearly separated dominant pair.

The example demonstrates that eliminating steady-state error does not automatically produce a fast response. Increasing K_i would reduce the slow-mode time scale in some cases but could also increase overshoot and reduce robustness. The final design must therefore be checked against actuator limits and transient specifications.

9. Illustrative Case: Position Control of a Six-Axis Robot Joint

Background

Consider an illustrative industrial robot joint driven by a servo motor and gearbox. The joint must track position commands accurately while carrying payloads from 2 to 8 kg. The encoder is high resolution, but its differentiated velocity estimate contains noise. The drive has current limits and a finite voltage supply.

Problem

A position-only proportional controller gives acceptable behavior at low payload but leaves a visible position error under gravity and friction. Increasing gain reduces error but excites a flexible gearbox mode. A direct PID controller also produces noisy commands and overshoot during large moves.

Analysis

The principal error sources are constant load torque, flexible transmission dynamics, actuator saturation during acceleration, and noisy velocity estimation. A single loop must simultaneously provide high low-frequency gain, adequate damping, and fast transient response. These requirements conflict.

Approach and solution

A cascade architecture is selected. The innermost current loop is tuned by the drive manufacturer and verified first. A velocity PI loop is then tuned with bandwidth well above the position-loop bandwidth. The outer position loop uses proportional action with a limited feedforward velocity term. Integral action is added only if residual static error remains after gravity compensation.

The position command is passed through a trajectory generator that limits velocity and acceleration. The velocity loop includes anti-windup, while the position loop has output limits corresponding to allowable velocity. Derivative action is avoided in the outer loop because the velocity loop already supplies damping and encoder differentiation is noisy.

The controller is tested at minimum and maximum payload, across the joint workspace, and under repeated reversals. Frequency-response measurements are used to ensure that the flexible mode remains outside the effective crossover region. Saturation tests verify that recovery after a large move does not produce excessive overshoot.

Outcome

The cascade design improves disturbance rejection and reduces interaction between position accuracy and motor-current dynamics. Trajectory limiting reduces saturation, and anti-windup shortens recovery after unavoidable limits. The design does not eliminate all error: gearbox backlash and friction still require mechanical improvement, compensation, or calibration.

Learning points

• Controller structure should reflect physical time scales.
• More gain is not a universal solution to tracking error.
• Feedforward, gravity compensation, and trajectory planning can reduce the burden on feedback.
• Saturation and sensor noise must be tested as primary design conditions, not as afterthoughts.

10. Design Trade-offs and Best Practices

Controller design is a trade-off among speed, accuracy, robustness, noise sensitivity, energy use, and implementation complexity. High bandwidth improves tracking of rapid commands but increases sensitivity to unmodeled dynamics and measurement noise. Integral action improves low-frequency accuracy but can create windup. Derivative or lead action adds damping but consumes phase margin when filtered and can amplify noise. Lag improves accuracy but commonly reduces response speed.

Recommended practice is to:

• define measurable requirements before tuning;
• use units consistently and document controller parameterization;
• identify actuator amplitude and rate limits;
• include sensor dynamics, computational delay, and filters in the loop model;
• tune from low gain toward the desired performance;
• use anti-windup whenever integral action and saturation coexist;
• validate across payload, temperature, friction, and supply-voltage conditions;
• log reference, measured output, error, unsaturated command, saturated command, and integrator state;
• provide independent emergency limits and safe shutdown behavior;
• protect software and communication interfaces against unauthorized changes in safety-critical machinery.

Modern tools may automate relay tests, frequency sweeps, parameter optimization, and gain scheduling. Machine-learning methods can assist model identification or adaptation, but they should not replace bounded commands, stability validation, fault detection, and human-defined safety constraints. In regulated industrial environments, traceability, verification, change control, and fail-safe behavior remain essential.

Chapter Summary

Classical controllers regulate mechanical and robotic systems using feedback and well-understood mathematical structures. On–off control is simple but limited in precision. Proportional control improves speed and reduces error, integral action removes steady-state error under suitable conditions, and derivative action adds damping but requires filtering. PID controllers combine these effects and must be tuned with attention to stability, bandwidth, actuator limits, and noise.

Lead compensation increases phase and can improve speed and stability margins. Lag compensation improves low-frequency accuracy at the cost of phase and bandwidth. Lead–lag designs combine both objectives. Cascade control separates physical time scales and is particularly effective in servo systems. Saturation, windup, noise, uncertainty, friction, compliance, and delay determine whether a theoretically satisfactory controller performs well in practice. Robust testing, anti-windup, trajectory limitation, and documented safety constraints are essential parts of professional controller design.

Key Terms

Feedback: Using measured output to influence a control command.

Reference: Desired value of the controlled variable.

Error: Difference between reference and measured output.

Plant: Physical system being controlled.

Transfer function: Input–output model in the Laplace domain under stated linear assumptions.

On–off controller: Controller that switches between discrete output states.

Proportional control: Control action proportional to instantaneous error.

Integral control: Control action based on accumulated error.

Derivative control: Control action based on error rate of change.

PID controller: Controller combining proportional, integral, and derivative actions.

Dead band: Error interval in which switching or action is deliberately withheld.

Overshoot: Amount by which a response exceeds its final value.

Settling time: Time required for a response to remain within a specified tolerance band.

Bandwidth: Frequency range over which a system tracks or responds effectively.

Phase margin: Frequency-domain measure related to robustness against instability.

Lead compensator: Compensator that contributes positive phase over a selected range.

Lag compensator: Compensator that increases low-frequency gain relative to high-frequency gain.

Cascade control: Nested control loops arranged from fast inner loops to slower outer loops.

Saturation: Limitation of actuator output or rate.

Integral windup: Excessive accumulation of integral state during actuator limitation.

Review Questions

Short Answer Questions

1. What is the principal purpose of feedback in a robotic control system?
2. Why can proportional control leave a nonzero steady-state error?
3. What benefit does integral action provide, and what practical problem can it cause?
4. Why is derivative action normally filtered?
5. State one condition under which cascade control is advantageous.

Long Answer Questions

1. Explain the operation of a PID controller and discuss the effect of each term on transient and steady-state performance.
2. Compare lead and lag compensation in terms of pole–zero placement, phase contribution, bandwidth, and application.
3. Describe a systematic procedure for tuning a PID controller in the presence of actuator limits and measurement noise.
4. Explain integral windup, derive the purpose of back-calculation anti-windup, and compare it with conditional integration.
5. Discuss how uncertainty, flexible mechanical modes, and delay influence the choice of controller bandwidth and stability margins.

Analytical / Application Questions

1. For a unity-feedback plant G(s) = 1/(s + 2) controlled by C(s) = K_p, derive the characteristic equation and determine the steady-state error to a unit step as a function of K_p.
2. A robot joint exhibits overshoot and slow recovery after its actuator saturates during a large move. Diagnose the likely control mechanism causing the behavior and propose at least three corrective actions.
3. Design a conceptual position–velocity cascade for a servo joint. Identify the controlled variable, actuator command, likely disturbance, sensor, and required relative bandwidth of the inner and outer loops.

Multiple Choice Questions with Answers

Question 1

Which controller action is primarily responsible for eliminating steady-state error to a constant reference in a stable type-0 system?

A. Proportional action

B. Integral action

C. Derivative action

D. Lead action

Correct Answer: B. Integral action

Question 2

For the standard second-order model, increasing the damping ratio ζ generally produces what effect?

A. Greater overshoot

B. Lower overshoot

C. Infinite bandwidth

D. More integral windup

Correct Answer: B. Lower overshoot

Question 3

In a practical PID controller, why is the derivative term filtered?

A. To eliminate actuator saturation

B. To increase integral gain

C. To limit amplification of high-frequency noise

D. To create a dead band

Correct Answer: C. To limit amplification of high-frequency noise

Question 4

Which compensator generally contributes positive phase near its design frequency?

A. Lead compensator

B. Lag compensator

C. Pure integrator

D. On–off controller

Correct Answer: A. Lead compensator

Question 5

For a standard lag compensator C(s) = K(1 + Ts)/(1 + βTs), which condition defines lag compensation?

A. β = 0

B. β < 0

C. 0 < β < 1

D. β > 1

Correct Answer: D. β > 1

Question 6

What is integral windup?

A. Sensor failure caused by high frequency

B. Accumulation of integral state while the actuator is saturated

C. A pole moving to infinity

D. Switching caused by a dead band

Correct Answer: B. Accumulation of integral state while the actuator is saturated

Question 7

In a cascade controller, which loop is normally tuned first?

A. The slowest outer loop

B. The reference generator

C. The fastest innermost loop

D. The safety shutdown loop only

Correct Answer: C. The fastest innermost loop

Question 8

For unity feedback with proportional control, plant gain K, and a unit-step reference, the stable type-0 steady-state error is:

A. K_pK

B. 1 + K_pK

C. 1/(1 + K_pK)

D. K_p/K

Correct Answer: C. 1/(1 + K_pK)

Question 9

Which action is most likely to produce derivative kick when the reference changes abruptly?

A. Derivative applied directly to the error

B. Integral applied to the measured output

C. Proportional action with saturation

D. Lag compensation at low frequency

Correct Answer: A. Derivative applied directly to the error

Question 10

What is a principal advantage of an inner current or torque loop in a robot servo?

A. It removes the need for position sensing

B. It can reject actuator-side disturbances before they strongly affect position

C. It guarantees zero overshoot in all conditions

D. It eliminates all mechanical flexibility

Correct Answer: B. It can reject actuator-side disturbances before they strongly affect position