Artificial Intelligence, Autonomy and Human–Robot Interaction

Introduction

Modern robots are no longer limited to executing fixed sequences of programmed motion. Industrial manipulators inspect products, mobile robots transport materials, collaborative robots share workspaces with people, and service robots interpret speech, gestures, and environmental context. These capabilities depend on artificial intelligence (AI), autonomous decision-making, and human–robot interaction (HRI).

A robot is autonomous when it can perceive its environment, interpret relevant information, select actions, and execute them with limited direct human intervention. Autonomy is not equivalent to intelligence in the human sense. It is a designed property with a defined operational scope, safety boundary, and responsibility structure. A robot may be highly autonomous in navigation but have no authority to modify its mission or operate outside approved conditions.

This chapter develops the principles needed to apply machine learning, reinforcement learning, behaviour architectures, explainable decision-making, collaborative robotics, teleoperation, and intuitive interfaces. Emphasis is placed on engineering judgement: selecting AI only where it improves performance, validating it against operational risks, and preserving meaningful human oversight.

Learning Objectives

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

1. Explain the relationship between perception, planning, control, autonomy, and HRI.
2. Distinguish supervised learning, unsupervised learning, reinforcement learning, and model-based programming.
3. Select suitable behaviour architectures for robotic tasks.
4. Formulate a basic reinforcement-learning problem using states, actions, rewards, and policies.
5. Design explainable and fail-safe autonomous decisions.
6. Compare collaborative operation, supervisory control, and teleoperation.
7. Evaluate interfaces using usability, workload, trust, and safety criteria.
8. Identify technical, ethical, legal, and organizational risks in autonomous robotics.

1. From Automation to Autonomy

Traditional automation follows a predetermined procedure. A programmable logic controller may activate a motor when a sensor reaches a threshold, while an industrial robot may repeat a trajectory generated offline. Such systems are predictable when the environment is controlled, but performance deteriorates when objects vary, people enter the workspace, or sensors become uncertain.

Autonomy adds adaptive decision-making. A useful abstraction is a closed loop:

Perception → World model → Decision → Action → Feedback

Perception converts sensor signals into estimates such as object position, human presence, or wheel speed. The world model stores relevant information, including uncertainty. Decision-making selects a task or motion. Action sends commands to actuators. Feedback measures the result and allows correction.

Autonomy should be specified by its operational design domain (ODD): the locations, objects, lighting, speeds, human roles, weather conditions, and failure states for which the robot has been validated. A warehouse vehicle may be autonomous on marked indoor routes but not on wet outdoor ground. Stating the ODD prevents the misleading claim that a system is simply “autonomous.”

Levels of human involvement

Human involvement can be organized as a continuum:

The appropriate mode depends on risk, communication quality, task complexity, and the consequences of error.

2. AI Methods for Robotics

2.1 Rule-based reasoning

Rule-based systems represent knowledge using conditions and actions, such as “if an obstacle is within the stopping distance, reduce speed.” They are transparent and easy to verify for limited situations. Their weaknesses are brittleness, rule conflicts, and poor performance in unanticipated environments.

Rules remain valuable for safety interlocks, operating limits, and supervisory logic even when machine learning is used for perception. A practical robot often combines learned components with deterministic safety constraints.

2.2 Supervised learning

In supervised learning, a model learns a mapping from labelled examples. For classification, the input may be an image and the output may be “acceptable part,” “damaged part,” or “unknown.” For regression, the output may be a continuous estimate such as object distance.

A dataset is normally divided into training, validation, and test sets. The test set must represent deployment conditions and must not influence model selection. Accuracy alone can be misleading when classes are imbalanced. For a binary classifier:

TP denotes true positives, FP false positives, and FN false negatives. Precision measures how many positive predictions are correct; recall measures how many actual positives are detected. In safety inspection, a high recall may be prioritized to avoid missing dangerous defects, although excessive false alarms create operational costs.

Machine-learning models may fail because of domain shift: the deployment environment differs from the training data. Changes in illumination, camera position, product design, dust, or operator clothing can reduce performance. Validation must therefore include representative variation and deliberate stress testing.

2.3 Unsupervised and self-supervised learning

Unsupervised learning identifies structure without manually supplied labels. Clustering can group similar vibration signatures, while dimensionality reduction can reveal operating states. These methods are useful for anomaly detection, but an anomaly is not automatically a fault. A newly introduced but harmless product may appear anomalous.

Self-supervised learning creates training signals from the data itself, such as predicting a missing image region or the next sensor value. It can reduce labelling effort, but deployment-specific validation remains necessary.

2.4 Reinforcement learning

Reinforcement learning (RL) models sequential decision-making. At time t, an agent observes a state sₜ, chooses an action aₜ according to a policy π, receives reward rₜ₊₁, and transitions to a new state sₜ₊₁.

The discounted return is:

Here Gₜ is the return from time t, r is reward, and γ is the discount factor, with 0 ≤ γ < 1 for continuing tasks. A larger γ gives greater importance to future rewards. Rewards must be dimensionally and operationally meaningful; they are design values rather than physical laws.

The optimal action-value function satisfies the Bellman optimality equation:

Q*(s,a) is the expected optimal return after taking action a in state s. The expectation accounts for stochastic transitions. RL is appropriate when actions have long-term consequences and explicit programming of every situation is difficult. It is risky when exploration can damage equipment or injure people. Training is therefore commonly performed in simulation, with safety constraints, demonstrations, offline data, or policy shields before limited real-world testing.

Worked example: discounted return

Suppose a robot receives rewards of 10, 4, and 2 over three successive steps, with γ = 0.8. The three-step return is:

The result is expressed in reward units. It is not a physical energy or monetary quantity unless the reward was explicitly defined that way. The calculation assumes that rewards after the third step are ignored.

3. Behaviour Architectures

A behaviour architecture determines how perception, goals, planning, and actions are organized.

3.1 Deliberative architectures

A deliberative robot constructs a model of the world, plans a sequence, and executes it. This supports long-horizon tasks such as route planning or assembly scheduling. However, planning can be computationally expensive and models may be inaccurate.

3.2 Reactive architectures

Reactive systems map current conditions directly to behaviours. A robot may stop when a proximity sensor detects a person or turn toward a visible target. Reactive control is fast and robust for local responses, but it may lack memory and fail to achieve complex goals.

3.3 Hybrid and layered architectures

A hybrid architecture combines a deliberative layer for goals and planning with reactive layers for immediate safety and obstacle avoidance. A common hierarchy is:

• mission layer: selects the task;
• task layer: sequences subtasks;
• skill layer: performs actions such as grasping or docking;
• safety layer: overrides unsafe commands.

The safety layer must have sufficient authority and low latency. A planner should never be able to command motion that violates a physical emergency stop or validated speed limit.

3.4 Behaviour trees and state machines

Finite-state machines represent discrete states and transitions, such as “search,” “approach,” “grasp,” and “verify.” They are understandable and suitable for structured tasks, but large systems can become difficult to maintain.

Behaviour trees use hierarchical nodes such as sequences, selectors, conditions, and actions. A sequence succeeds only when its required child actions succeed; a selector tries alternatives until one succeeds. Behaviour trees support modularity and explicit fallback behaviour, making them useful for service and industrial robots.

4. Perception, Uncertainty and Decision-Making

Sensors provide incomplete and noisy evidence. A camera may be affected by glare, a lidar may miss dark surfaces, and an encoder may drift. Autonomous decisions should therefore distinguish a measurement from a belief about the world.

If a robot estimates that an object is at position x̂ with standard uncertainty σ, a decision should consider both values. A low-confidence grasp may trigger another observation rather than immediate motion. Confidence thresholds must be selected using operational consequences, not arbitrary percentages.

A robust decision pipeline includes:

1. sensor validation and plausibility checks;
2. estimation of object or human state;
3. confidence assessment;
4. selection among action, observation, slowdown, and stop;
5. logging of the evidence and decision;
6. monitoring of the outcome.

An explainable decision does not require exposing every internal neural-network weight. It should provide a useful account of the evidence, objective, constraint, and uncertainty. For example: “The robot slowed because a person was detected within the shared-zone boundary; detection confidence was 0.91, and the speed limit for this zone is 0.25 m/s.”

Explainability has several audiences. An operator needs an immediate reason and recovery instruction. A maintenance engineer needs sensor and fault information. A safety assessor needs traceability to requirements. A regulator or investigator may need time-stamped logs and model versions.

5. Human–Robot Interaction

HRI concerns the technical and social relationship between people and robots. It includes physical interaction, communication, shared planning, trust, workload, and responsibility.

5.1 Collaborative robotics

A collaborative robot is designed to operate in proximity to people under specified safeguarding conditions. Collaboration is not created merely by removing a fence. Risk assessment must consider robot mass, speed, payload, tool geometry, reachable workspace, pinch points, unexpected restart, and the task performed by the person.

Common collaborative strategies include:

• safety-rated monitored stop;
• hand guiding;
• speed and separation monitoring;
• power and force limiting.

These strategies have different assumptions and limitations. Force limiting may reduce injury severity but does not eliminate risks from sharp tools, entrapment, or high-energy payloads. A risk assessment should examine the complete application rather than the robot arm alone.

5.2 Communication channels

Robots communicate through:

• visual signals such as lights, displays, and projected paths;
• auditory signals such as tones and spoken messages;
• haptic signals such as vibration or force feedback;
• physical motion, including intentional gestures;
• natural-language interaction.

Each channel has limitations. Audible alarms may be masked by factory noise, while visual displays may be ignored when an operator is looking at a workpiece. Safety-critical information should be redundant or delivered through the channel most reliably perceived in the context.

5.3 Trust and appropriate reliance

Trust should be calibrated to actual capability. A robot that appears confident while frequently failing can cause misuse; a robot that constantly requests confirmation can create automation neglect. Interfaces should communicate confidence, task status, limitations, and reasons for intervention without overwhelming the operator.

5.4 Teleoperation and shared control

Teleoperation separates the operator from the physical robot. A teleoperator uses a command device and receives visual, auditory, or haptic feedback. Important design variables include latency, camera viewpoint, field of view, control mapping, bandwidth, and workload.

Shared control assists the operator while preserving human intent. For example, the operator may command motion toward a shelf while the robot prevents collision with the shelf. The assistance policy must be predictable; unexpected intervention can reduce trust and may cause the operator to fight the system.

6. Designing Intuitive Interfaces

An effective interface follows the principle that the user should understand what the robot is doing, why it is doing it, and how to influence it. Good design begins with task analysis rather than selecting a fashionable technology.

Important interface properties include:

• observability: the user can perceive robot state and progress;
• controllability: the user can start, pause, redirect, or stop the robot;
• consistency: similar commands produce similar outcomes;
• feedback: commands and errors are acknowledged;
• recoverability: the user can correct mistakes without unsafe improvisation;
• accessibility: the interface accommodates language, sensory, and physical differences.

Natural-language interfaces can make robots easier to command, but ambiguity is dangerous. A robot should confirm unclear high-impact commands, constrain language to the validated task domain, and distinguish planning from execution. For example, “move the box” should not directly trigger motion if multiple boxes or destinations are possible.

Gesture interfaces require attention to accidental gestures, occlusion, lighting, and cultural interpretation. Wearable interfaces may improve control but introduce privacy and comfort concerns. Touchscreens are efficient for structured choices but may be unsuitable when the operator wears gloves.

7. Safety, Ethics and Governance

Autonomous robots distribute responsibility among designers, manufacturers, deployers, operators, and maintainers. A technical system cannot resolve every responsibility question, but engineering practice can make responsibility clearer.

Core governance practices include:

1. define the intended use and prohibited use;
2. document assumptions and operating limits;
3. perform hazard identification and risk reduction;
4. validate normal and abnormal conditions;
5. maintain version-controlled software and models;
6. log commands, sensor states, decisions, and overrides;
7. provide emergency stop and safe-state behaviour;
8. train operators and define authority boundaries;
9. monitor incidents, near misses, and model drift;
10. review the system after changes in environment or task.

Ethical issues include privacy from cameras and microphones, bias in recognition systems, unequal distribution of workplace benefits and risks, worker surveillance, and the dignity of human participants. A robot should collect only necessary data, protect access to it, and communicate when recording occurs.

Autonomous systems should be designed according to applicable machinery safety, electrical safety, cybersecurity, data-protection, and occupational-health requirements. Standards and regulations vary by jurisdiction and application; therefore, engineers must consult current authoritative requirements rather than relying on informal summaries.

Cybersecurity is part of physical safety. A compromised navigation system, stolen credential, or manipulated sensor can create hazardous motion. Network segmentation, authentication, secure updates, least-privilege access, anomaly monitoring, and recovery procedures are essential.

8. Illustrative Case: Collaborative Inspection Cell

Background

A manufacturer wants a robot to present small metal components to a human inspector. The robot retrieves parts from a tray, positions each part under a camera, and places accepted or rejected parts into separate bins.

Problem

Parts vary in orientation, the inspection camera sometimes produces low-confidence results, and the operator must be able to intervene when a part is damaged or obstructed. The robot will work beside a person, so unexpected motion and pinch points are significant hazards.

Analysis

A purely scripted system is unsuitable because part orientation varies. A fully autonomous system is also inappropriate because low-confidence defects require human judgement. The task therefore needs learned perception, deterministic motion limits, a behaviour tree, and supervisory human control.

Approach and solution

A vision model classifies part condition and estimates orientation. The model is trained on representative images containing variation in lighting, surface finish, and orientation. A validation set is held aside, and low-confidence predictions are routed to the inspector.

The behaviour tree follows this logic:

1. verify that the cell is clear;
2. detect and estimate a part;
3. if confidence is low, request inspection;
4. if confidence is adequate, plan a collision-checked approach;
5. move at a validated collaborative speed;
6. present the part and display the reason for classification;
7. accept operator correction;
8. place the part and record the event.

The safety system independently monitors the shared workspace. If a person enters the separation boundary, the robot slows or stops according to the validated safeguarding strategy. The operator has a clearly visible stop control and can select “reinspect,” “accept,” or “reject.”

Outcome

The combined architecture preserves automation for repetitive handling while assigning ambiguous decisions to a human. Logs record image identifier, model version, confidence, operator action, and robot state. Performance is evaluated using inspection recall, false-rejection rate, cycle time, intervention frequency, and near-miss reports—not cycle time alone.

Learning points

The case demonstrates that autonomy should be allocated selectively. Machine learning is valuable for variable visual perception, while deterministic constraints are preferable for safety. Human intervention is not a failure of automation when it is intentionally designed as part of the operating concept. Explainability, logging, and fallback behaviours make the system easier to validate and improve.

9. Evaluation and Best Practices

A robotics AI project should establish measurable requirements before model selection. Useful measures include task success rate, collision rate, false-positive and false-negative rates, intervention frequency, latency, energy use, operator workload, and recovery time.

Testing should progress from simulation and recorded data to controlled physical trials and finally supervised deployment. Simulation can expose rare situations economically, but it may omit friction, sensor artefacts, human unpredictability, and mechanical wear. Real-world testing must therefore include uncertainty and degraded conditions.

Best practices are to:

• use the simplest method that meets the requirement;
• separate perception uncertainty from safety authority;
• keep safety functions independent of experimental learning components;
• design explicit fallback and recovery behaviours;
• test distribution shift and adversarial conditions;
• communicate capability boundaries to users;
• involve operators in interface design;
• treat model updates as engineering changes requiring revalidation;
• monitor performance after deployment;
• preserve human authority for high-consequence decisions.

Contemporary Developments

Foundation models and multimodal systems are increasingly used to interpret language, images, and task descriptions. They may improve flexible interaction, but their outputs can be uncertain, difficult to verify, and vulnerable to ambiguous instructions. A safe architecture places such models above validated planning and control layers rather than allowing unrestricted direct actuator access.

Digital twins support simulation, predictive maintenance, and operator training. Edge computing reduces communication latency and may protect sensitive data, whereas cloud systems provide scalable computation and fleet-level learning. Continual learning offers adaptation but creates risks if the robot changes behaviour without controlled approval. These technologies strengthen robotics only when accompanied by monitoring, version control, and safety validation.

Chapter Summary

AI enables robots to interpret variable environments and select actions, but autonomy must be bounded by an operational design domain and safety requirements. Supervised learning is useful for labelled perception tasks; unsupervised and self-supervised methods discover structure; reinforcement learning addresses sequential decisions but requires carefully designed rewards and safe training. Behaviour architectures combine reactive responses, planning, task sequencing, and safety overrides.

Human–robot interaction includes physical collaboration, teleoperation, shared control, communication, trust, workload, and interface design. Intuitive systems make robot state, intentions, uncertainty, and intervention methods clear. Explainable autonomy records and communicates relevant evidence, objectives, constraints, and confidence. Ethical governance requires privacy protection, cybersecurity, accountability, operator training, model validation, and continuous monitoring. The most effective contemporary systems are hybrid: learned where variability is high, deterministic where safety and verification are essential, and human-centred where judgement and responsibility remain important.

Key Terms

Autonomy: The capability of a system to perform defined tasks with limited direct human control.

Operational design domain: The conditions under which a robotic system is designed and validated to operate.

Supervised learning: Learning from labelled input–output examples.

Reinforcement learning: Learning sequential decisions through interaction and reward signals.

Policy: A rule or model that selects actions from states or observations.

State: Information representing the relevant condition of a system and environment.

Reward: A numerical signal used to evaluate an action or outcome in reinforcement learning.

Behaviour tree: A hierarchical structure for sequencing, selecting, and recovering behaviours.

Finite-state machine: A model consisting of discrete states and defined transitions.

Reactive control: Action selection based primarily on current sensor conditions.

Deliberative control: Planning based on an internal model of tasks or the environment.

Human–robot interaction: The study and design of communication and cooperation between people and robots.

Teleoperation: Remote control of a robot using command and feedback channels.

Shared control: Human control assisted or constrained by autonomous functions.

Explainability: The ability to provide understandable reasons for system outputs or actions.

Domain shift: A difference between training conditions and deployment conditions.

Collaborative robot: A robot designed for specified operation in proximity to people.

Operational technology cybersecurity: Protection of control systems, networks, and devices whose compromise may affect physical processes.

Model drift: Degradation or change in model performance as data or operating conditions change.

Safety case: A structured argument and evidence that a system is acceptably safe for its intended use.

Review Questions

Short Answer Questions

1. Define autonomy and explain why it must be associated with an operational design domain.
2. Distinguish supervised learning from reinforcement learning.
3. What is the purpose of a safety layer in a hybrid behaviour architecture?
4. Define teleoperation and shared control.
5. List four properties of an effective human–robot interface.

Long Answer Questions

1. Explain the perception–world model–decision–action–feedback loop and discuss sources of uncertainty at each stage.
2. Compare deliberative, reactive, and hybrid behaviour architectures, including an application for each.
3. Explain how explainability can support operators, maintenance engineers, and safety investigators.
4. Discuss the principal hazards and safeguarding considerations in collaborative robotics.
5. Describe a validation strategy for a robot using machine learning in a changing industrial environment.

Analytical / Application Questions

1. Design a behaviour architecture for an autonomous warehouse robot that must transport materials while sharing aisles with people. Identify the role of learning, planning, reactive safety, and human intervention.
2. A vision classifier has high accuracy on laboratory images but frequently misses defects under factory lighting. Analyse likely causes and propose a test and improvement plan.
3. Design an interface for a teleoperated robot used in hazardous inspection. Justify your choices of feedback channels, control devices, alerts, logging, and emergency procedures.

Multiple Choice Questions with Answers

Question 1

Which statement best describes an operational design domain?

A. The maximum computing speed of a robot

B. The conditions for which a robot has been designed and validated

C. The number of actions in a reinforcement-learning policy

D. The physical dimensions of a robot cell

Correct Answer: B. The conditions for which a robot has been designed and validated

Question 2

In supervised learning, the model is trained primarily using:

A. Labelled examples

B. Random motor commands only

C. Unrestricted human conversation

D. Undiscounted rewards without observations

Correct Answer: A. Labelled examples

Question 3

For γ = 0.8 and rewards 10, 4, and 2, the three-step discounted return is:

A. 12.00

B. 13.20

C. 14.48

D. 16.00

Correct Answer: C. 14.48

Question 4

Which architecture is most suitable for combining long-term planning with immediate obstacle avoidance?

A. Purely reactive architecture

B. Purely mechanical architecture

C. Hybrid or layered architecture

D. Unsupervised architecture

Correct Answer: C. Hybrid or layered architecture

Question 5

Recall is calculated as:

A. TP ÷ (TP + FP)

B. TP ÷ (TP + FN)

C. TN ÷ (TN + FP)

D. FP ÷ (TP + FP)

Correct Answer: B. TP ÷ (TP + FN)

Question 6

What is a principal purpose of explainability in an autonomous robot?

A. To guarantee that the model is always correct

B. To eliminate the need for testing

C. To communicate relevant reasons, evidence, and uncertainty

D. To increase motor torque

Correct Answer: C. To communicate relevant reasons, evidence, and uncertainty

Question 7

Shared control means that:

A. Two robots always perform identical motions

B. A human commands the robot while autonomous assistance constrains or supports operation

C. The robot operates without any human authority

D. The robot is controlled only by a remote camera

Correct Answer: B. A human commands the robot while autonomous assistance constrains or supports operation

Question 8

Which issue is an example of domain shift?

A. A robot receiving a new battery

B. Deployment lighting differing substantially from training lighting

C. A safety stop being pressed

D. A planner completing its route

Correct Answer: B. Deployment lighting differing substantially from training lighting

Question 9

Why is unrestricted direct actuator access by a general-purpose language model risky?

A. Language models cannot produce text

B. The model may generate uncertain or ambiguous commands without validated safety constraints

C. Actuators do not require safety systems

D. It prevents all forms of autonomy

Correct Answer: B. The model may generate uncertain or ambiguous commands without validated safety constraints

Question 10

Which practice best supports responsible deployment of learning-enabled robots?

A. Disable logging to protect performance

B. Avoid testing outside the laboratory

C. Version models, validate updates, and monitor deployment performance

D. Allow the system to change objectives without approval

Correct Answer: C. Version models, validate updates, and monitor deployment performance