Intelligent Control, Robotics and Autonomous Systems

Introduction

Mechanical systems are increasingly connected, instrumented and software-defined. A machine tool may combine servo drives, programmable logic controllers (PLCs), industrial networks, cloud analytics and machine-learning models. A mobile robot may perceive its environment, estimate its position, plan a path and control its motion while operating near people. These systems are cyber-physical systems: physical processes are influenced by computation, communication and data.

Intelligent control does not replace control engineering. Rather, it extends conventional feedback control with estimation, adaptation, prediction, optimization and learning. The engineer must still establish a safe plant model, select sensors and actuators, guarantee acceptable dynamic performance and manage uncertainty. AI is valuable only when it is integrated within a sound control architecture.

This chapter examines feedback control, PLC automation, industrial and collaborative robots, autonomous mobile systems, adaptive and predictive control, reinforcement learning, communication latency and functional safety.

Learning Objectives

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

1. Explain how feedback, estimation and AI are combined in cyber-physical mechanical systems.
2. Distinguish PLC, industrial-robot, collaborative-robot and autonomous-mobile-system architectures.
3. Formulate basic state-space, adaptive-control and model-predictive-control relationships.
4. Analyze the effects of sampling, communication delay and uncertainty on stability and performance.
5. Identify appropriate uses and limitations of machine learning and reinforcement learning in control.
6. Design a safety-oriented architecture for an intelligent automated cell.

1. Feedback Control as the Foundation

A feedback controller measures a controlled variable, compares it with a desired reference and commands an actuator. For a single-input, single-output system, the error is

where r(t) is the reference, y(t) is the measured output and e(t) is the error. A proportional-integral-derivative controller is commonly written as

Here u(t) is the control signal, and K_p, K_i and K_d are proportional, integral and derivative gains. The expression assumes a sufficiently smooth measured signal and a controller whose output and actuator remain within allowable limits. In practical digital systems, derivative action is filtered and integral action is protected against wind-up.

A simplified negative-feedback loop has closed-loop transfer function

where G(s) is the plant, C(s) the controller and H(s) the measurement path. Stability depends on the poles of the closed-loop system, not merely on the quality of a prediction model. A learned component should therefore normally operate inside a supervisory or constrained control layer rather than directly bypassing all safety-critical feedback.

State-Space Representation

For multi-variable mechanical systems, state-space notation is useful:

The vector x contains states such as position and velocity; u contains actuator inputs; y contains measured outputs; and A, B, C and D are system matrices of compatible dimensions. This linear model applies locally or globally only when linearization is justified. Nonlinear systems may instead be represented by ẋ = f(x,u) and y = h(x,u).

Sensors rarely measure every state. An observer or Kalman filter combines a model and noisy measurements to estimate unmeasured states. For a discrete linear model,

where w_k and v_k represent process and measurement noise. A Kalman filter is appropriate when the model is approximately linear and noise assumptions are credible; it is not a universal solution for severe nonlinearities or biased sensors.

2. PLC Automation and Industrial Networks

A PLC executes a cyclic scan: input acquisition, logic execution, output update and diagnostic tasks. Ladder logic, function-block diagrams and structured text support deterministic sequencing, interlocking and timing. Typical PLC applications include conveyor coordination, hydraulic sequencing, pressing, packaging and machine-tool auxiliaries.

AI can enhance a PLC system through condition monitoring, anomaly detection, quality prediction and set-point optimization. The PLC should retain deterministic interlocks, emergency-stop logic and limit protection. A non-deterministic cloud model should not be the sole mechanism preventing a dangerous actuator command.

A practical architecture separates three layers:

Industrial Ethernet, fieldbuses and wireless systems connect these layers. Engineering decisions must consider determinism, jitter, packet loss, cybersecurity and time synchronization. Edge computing is often preferred for fast decisions because it reduces dependence on wide-area networks.

3. Industrial Robots and Collaborative Robots

An industrial robot is a programmable manipulator consisting of links, joints, actuators, sensors and a controller. Its configuration is represented by joint coordinates q. Forward kinematics maps joint coordinates to end-effector pose:

The Jacobian relates joint velocity to end-effector velocity:

Near a singular configuration, J(q) loses rank or becomes poorly conditioned, so small Cartesian motions may require large joint velocities. Singularity avoidance, payload limits and collision checking are essential in programming robot trajectories.

Industrial robots usually operate in segregated cells with guarding, scanners or interlocked gates. Collaborative robots, or cobots, are designed to share a workspace under specified conditions. Collaboration does not mean that every contact is harmless. Risk assessment must consider speed, force, tool geometry, pinch points, payload, human behavior and foreseeable misuse. Power-and-force limiting, speed-and-separation monitoring, hand guiding and safety-rated monitored stops are examples of collaborative strategies.

Machine vision and AI can support grasp selection, bin picking, surface inspection and human-intent estimation. However, perception uncertainty must be managed with confidence thresholds, fallback actions and safe stopping. A vision model that misclassifies a person as background is a safety concern, not merely a prediction error.

4. Autonomous Mobile Systems

An autonomous mobile robot (AMR) normally performs four coupled functions:

1. Perception: detecting obstacles, landmarks and humans.
2. Localization and mapping: estimating pose and, when necessary, constructing a map.
3. Planning: selecting a collision-free route and feasible trajectory.
4. Control: tracking the trajectory while respecting vehicle dynamics.

For a planar differential-drive robot,

where x,y are planar coordinates, θ is heading, v is forward speed and ω is angular speed. The model assumes ideal wheel rolling and negligible lateral slip. Real robots require calibration and compensation for wheel slip, backlash and uneven floors.

Localization may combine wheel odometry, inertial measurements, laser scanning and cameras. Simultaneous localization and mapping (SLAM) estimates both robot pose and environmental structure. Path planning algorithms such as graph search are useful for discrete maps, whereas sampling-based methods can handle high-dimensional configuration spaces. A local controller must react to moving obstacles even when the global route is unchanged.

5. Adaptive and Predictive Control

A fixed controller can perform poorly when payload, friction, temperature or cutting conditions vary. Adaptive control updates parameters during operation. In indirect adaptive control, the system is identified first and the controller is redesigned from the estimated model. In direct adaptive control, controller parameters are adjusted without explicitly identifying every plant parameter.

A simple parameter-estimation model is

where φ_k is a regressor vector, θ an unknown parameter vector and ε_k the modeling error. A recursive least-squares update can be used when the relationship is approximately linear in parameters, persistent excitation is sufficient and noise assumptions are reasonable. Adaptation must be bounded; projection, dead zones and parameter limits help prevent drift caused by noise or unmodeled dynamics.

Model predictive control (MPC) repeatedly solves a finite-horizon optimization problem:

subject to the plant model and constraints such as actuator limits, velocity limits and collision-avoidance regions. Here N is the prediction horizon; Q, R and P are weighting matrices, normally positive semidefinite or positive definite as appropriate; r_i is the reference; and x_N is the terminal state. Only the first computed input is applied, after which measurements are obtained and the optimization is repeated. MPC is powerful because it handles constraints explicitly, but computational cost, model mismatch and feasibility must be addressed.

Worked Example: Sampling and Delay

Suppose a position-control loop has a sampling period of 10 ms and communication introduces 25 ms of delay. The command is therefore based on information that may be approximately 35 ms old, excluding computation and actuator delay. If the mechanical plant has a dominant time constant of 20 ms, the delay is larger than that time constant and can substantially reduce phase margin. The correct engineering response is not simply to increase gain. The designer should measure the complete delay, reduce network and computation latency, lower bandwidth if necessary, add prediction or redesign the architecture, and verify stability experimentally and analytically.

6. Reinforcement Learning in Mechanical Control

Reinforcement learning (RL) models sequential decision-making. At time step t, an agent observes state s_t, selects action a_t, receives reward r_t and transitions to s_t+1. A discounted return is

where 0 ≤ γ < 1 is the discount factor. The objective is to learn a policy that maximizes expected return. In mechanical engineering, RL has been investigated for energy management, robot manipulation, gait generation and motion planning.

Direct online exploration on a high-energy machine is risky. Safer approaches include simulation training, domain randomization, offline learning from validated data, action shielding, constraint filters and a conventional fallback controller. The simulation-to-reality gap arises from inaccurate friction, contacts, sensor noise, delays and unmodeled wear. RL is therefore best treated as a candidate policy generator or optimizer unless rigorous safety and performance evidence is available.

7. Stability, Latency and Functional Safety

A stable nominal controller may become unsafe when sensing, computation and actuation are separated by a network. Sampling and delay add phase lag. Variable delay, known as jitter, makes worst-case analysis more important than average latency. Relevant measures include sampling period, end-to-end delay, jitter, packet-loss rate, watchdog timeout and maximum stopping distance.

Functional safety is the capability of a control system to reduce risk by performing safety functions correctly. Examples include emergency stop, guard-door monitoring, safe torque off, speed limitation and safe position monitoring. Safety-related functions should be designed according to applicable standards and validated through hazard analysis, fault analysis, testing and documented lifecycle processes. Exact performance-level or safety-integrity targets depend on the machinery, jurisdiction and risk assessment; they must not be guessed from a generic AI model accuracy.

A useful principle is separation of authority. The AI layer may recommend a trajectory or detect an anomaly, while an independently engineered safety layer limits speed, workspace, torque and stopping behavior. Cybersecurity is also part of safety: a compromised sensor, PLC or update mechanism can create physical hazards. Secure authentication, network segmentation, signed software, least privilege, logging and controlled updates are therefore engineering requirements.

8. Illustrative Case: Intelligent Robotic Palletizing Cell

Background

A distribution facility uses a six-axis robot to palletize boxes from a conveyor. Box dimensions vary, and a camera identifies orientation and location. The cell includes a PLC, robot controller, safety scanner, conveyor drive and edge computer.

Problem

Fixed recipes caused frequent pauses when box locations changed. Management proposed allowing a neural network to command robot joint positions directly. This approach could improve flexibility but would create unacceptable uncertainty near workers and equipment.

Analysis

The engineering team separated the problem into perception, planning, control and safety. The vision model estimated box pose and supplied a confidence score. A calibrated geometric planner generated grasp and palletizing poses. The robot controller executed a collision-checked trajectory. The PLC supervised sequence logic, presence sensors and conveyor interlocks. A safety scanner initiated a safety-rated monitored stop independently of the AI computer.

Approach and Solution

The edge computer performed inference locally to reduce latency. If confidence fell below a validated threshold, the system requested a re-scan or human intervention. The planner rejected poses outside the workspace or beyond payload limits. A watchdog stopped the task if communication was lost. Production data were logged for drift monitoring, but model updates were tested offline and approved before deployment.

Outcome

The cell became more tolerant of variable box placement while retaining deterministic motion and independent protective functions. The main benefit came from combining AI perception with conventional planning and control, not from replacing the robot controller with an end-to-end neural network.

Learning Points

The case illustrates defense in depth, bounded AI authority, local computation, human override, traceability and the importance of validating the complete cyber-physical system rather than only the classifier.

9. Professional Design Method

A systematic design workflow is:

1. Define the mechanical task, performance metrics and operating envelope.
2. Identify hazards, foreseeable misuse and required safety functions.
3. Select sensors with suitable range, bandwidth, accuracy and environmental rating.
4. Establish the deterministic control loop before adding AI.
5. Characterize latency, jitter, failure modes and data quality.
6. Add AI only where it offers measurable value, such as prediction or perception.
7. Constrain AI outputs using physical limits, safety monitors and fallback modes.
8. Test nominal, boundary, fault, cyberattack and recovery conditions.
9. Monitor drift, calibration, maintenance condition and human interaction after deployment.

Important performance measures include tracking error, settling time, overshoot, energy use, throughput, collision rate, false alarms, missed detections, availability and safe-stop performance. AI accuracy alone is insufficient because a high average accuracy can conceal rare but dangerous failures.

Contemporary Developments and Limitations

Digital twins support commissioning and what-if analysis, but their usefulness depends on model fidelity and validated data. Edge AI reduces latency and protects some data from leaving the factory. Human-aware robotics uses vision, force sensing and behavioral prediction to improve cooperation. Foundation models may simplify natural-language programming, yet generated commands require access control, verification and testing.

Persistent limitations include distribution shift, sensor degradation, unmodeled contact dynamics, adversarial inputs, data bias, explainability difficulties and maintenance burden. Autonomous systems should expose uncertainty and fail safely rather than present unwarranted confidence. Engineers remain accountable for design decisions, documentation and compliance.

Chapter Summary

Intelligent control integrates conventional feedback, estimation, optimization and learning within a cyber-physical architecture. PLCs provide deterministic sequencing and interlocking; industrial robots provide repeatable manipulation; cobots require task-specific risk assessment; and AMRs combine perception, localization, planning and motion control. Adaptive control responds to changing plant parameters, while MPC handles predictions and constraints. Reinforcement learning can discover useful policies but requires simulation, constraint protection and rigorous validation.

Sampling, delay, jitter and packet loss influence stability and must be considered as part of the control loop. Functional safety, cybersecurity, independent protective layers and controlled software updates are essential. The strongest engineering solutions assign AI limited, monitorable authority and retain validated feedback and safety mechanisms beneath it.

Key Terms

Cyber-physical system: A system integrating physical machinery with computation, communication and sensing.

Feedback control: Control based on measured output and the difference between reference and output.

PLC: Programmable logic controller used for deterministic industrial automation.

Edge computing: Processing data near the physical equipment rather than exclusively in a remote cloud.

State estimation: Inferring internal system states from models and measurements.

Observer: An algorithm that estimates unmeasured states.

Robot kinematics: The relationship between joint motion and end-effector position or velocity.

Jacobian: Matrix mapping joint velocities to Cartesian velocities.

Collaborative robot: Robot designed for specified forms of shared human workspace operation.

AMR: Autonomous mobile robot capable of perception, navigation and motion control.

SLAM: Simultaneous localization and mapping.

Adaptive control: Control that updates parameters as plant behavior changes.

MPC: Model predictive control using repeated finite-horizon optimization.

Reinforcement learning: Learning action policies through interaction and reward.

Latency: Time between sensing, computation, communication and actuation.

Jitter: Variation in timing or delay.

Functional safety: Risk reduction achieved through correctly performing safety functions.

Fail-safe state: Condition designed to reduce hazard when a fault occurs.

Digital twin: A computational representation linked to a physical asset or process.

Review Questions

Short Answer Questions

1. Why should a learned model normally not replace the complete safety layer of a machine?
2. What functions are performed by perception, localization, planning and control in an AMR?
3. Define latency and jitter in a networked control system.
4. State two differences between an industrial robot cell and a collaborative robot application.
5. Why is edge computing often useful in intelligent control?

Long Answer Questions

1. Explain the architecture of a PLC-based intelligent manufacturing cell and identify suitable roles for AI.
2. Derive the meaning of the state-space model and discuss the assumptions of a linear representation.
3. Compare adaptive control, MPC and reinforcement learning for a variable-payload robot.
4. Explain how communication delay can affect stability and describe engineering mitigation strategies.
5. Discuss functional safety, cybersecurity and human factors in a collaborative robotic system.

Analytical / Application Questions

1. Design a layered control and safety architecture for an AMR operating in a warehouse shared with pedestrians. Specify sensors, controllers, fallback behavior and performance metrics.
2. A robot vision system has high average classification accuracy but occasionally misidentifies a person as a box. Analyze the risk and propose technical and organizational controls.
3. For the palletizing case, decide which functions should run in the PLC, robot controller, safety controller and edge computer. Justify the allocation using determinism, latency, safety and maintainability.

Multiple Choice Questions with Answers

Question 1

Which component should normally retain deterministic emergency-stop and interlocking functions?

A. Cloud analytics platform

B. Safety controller or validated PLC logic

C. Unconstrained reinforcement-learning policy

D. Production database

Correct Answer: B. Safety controller or validated PLC logic

Question 2

In the planar differential-drive model, what does ω represent?

A. Forward speed

B. Wheel radius

C. Angular speed

D. Planar x-coordinate

Correct Answer: C. Angular speed

Question 3

What is the principal advantage of MPC identified in the chapter?

A. It eliminates the need for a plant model

B. It handles predictions and constraints explicitly

C. It guarantees safety without validation

D. It requires no computation during operation

Correct Answer: B. It handles predictions and constraints explicitly

Question 4

A Jacobian in robot kinematics maps:

A. Joint velocities to end-effector velocities

B. Temperature to pressure only

C. Images directly to safety ratings

D. Rewards to discount factors

Correct Answer: A. Joint velocities to end-effector velocities

Question 5

Which condition is especially important before applying reinforcement learning online to a high-energy machine?

A. Unlimited exploration

B. Removal of all fallback controllers

C. Constraint protection and validated fallback behavior

D. Ignoring actuator limits

Correct Answer: C. Constraint protection and validated fallback behavior

Question 6

What is jitter?

A. A fixed reference position

B. Variation in timing or delay

C. A robot’s payload rating

D. A type of mechanical bearing

Correct Answer: B. Variation in timing or delay

Question 7

Why can communication delay reduce control-loop stability?

A. It adds phase lag to the loop

B. It always increases sensor accuracy

C. It removes actuator saturation

D. It guarantees zero tracking error

Correct Answer: A. It adds phase lag to the loop

Question 8

Which task belongs primarily to AMR perception?

A. Detecting obstacles and landmarks

B. Computing motor winding temperature only

C. Writing PLC ladder logic

D. Approving software changes

Correct Answer: A. Detecting obstacles and landmarks

Question 9

In the illustrative palletizing cell, why was AI not allowed to command unconstrained joint positions directly?

A. AI cannot process images

B. The approach would create unacceptable uncertainty near people and equipment

C. Robot controllers cannot execute trajectories

D. Conveyor systems never require sensors

Correct Answer: B. The approach would create unacceptable uncertainty near people and equipment

Question 10

What does functional safety concern?

A. The color of a robot enclosure

B. The capability of safety functions to reduce risk correctly

C. Maximizing cloud storage

D. Increasing model complexity regardless of risk

Correct Answer: B. The capability of safety functions to reduce risk correctly