Human–Robot Interaction: An Interactive Introduction

Physical and social interaction

A robot picking boxes in an empty warehouse only has to handle the physics of its own body and the environment. But as soon as it brings that box to a warehouse worker, or carries a coffee to a café customer, the robot has to think about social things on top of the physical motion: where the people are, how to approach them safely and appropriately, and how to follow social rules such as answering "thank you" with "you're welcome". These rules are obvious to people and entirely unknown to the robot, so the designer has to supply every one of them. That is what makes the questions of HRI different from the questions of robotics alone.

What embodiment changes

What sets HRI apart from other computing technologies is that the robot has a body: embodiment. The body constrains how the robot can sense and act, but it is also an affordance that makes people treat the robot like a person. If the robot has eyes, people assume it can see; if it has a mouth, that it can talk. People bring their experience of human interaction straight over to the robot, and when the robot falls short of those expectations, disappointment follows (chapters 2 and 8).

Where the disciplines meet

HRI is a multidisciplinary field in which engineers, psychologists, designers, anthropologists, sociologists and philosophers work together. The textbook (section 2.2) summarises three walls between the groups: how knowledge is expressed (engineers and scientists publish it explicitly; designers embed it in the artefact and in intuition), how reality is viewed (engineers and designers want to change the world; scientists want to understand it), and the main concern (technology for engineers; people for designers and scientists). Knowing these differences in advance makes collaboration easier.

A short history

The term "social robot" was first used for robots in a 1978 magazine article. The behaviour-based robots of the 1990s produced a lifelike impression from simple reactive rules alone, changing ideas about how apparent intelligence arises. MIT's Kismet (1997) was an early social robot that moved its eyes, eyebrows, lips and neck in response to people's voices and movements. With only a handful of social drives, it used big eyes and an exaggerated face (the baby schema) to project a surprisingly social presence. Commercial platforms such as Nao (2006–), Pepper, Keepon and Paro then allowed many labs to replicate and compare studies on the same robot.

Inside-out and outside-in

Robots are usually built by engineers first and tested by social scientists later: bolt sensors and an arm onto a mobile platform such as a TurtleBot, and add a shell if there is time. The textbook (ch. 4) calls this stitching-together of existing technology the "Frankenstein approach" and considers it poor practice, because it ignores the user and the context. The opposite direction is outside-in design: start from the interaction the robot is meant to have, decide on the appearance and behaviour, and fill in the technology. ASIMO was made small so as not to be threatening; Pepper was given a waist that can bow to greet customers in Japanese shops. The seal robot Paro started as a cat, but people knew real cats too well and compared it unfavourably, so it became a seal. It also had wheels, which were removed because its main users, older adults with limited mobility, had no need for a robot that roams.

Choices of form

• Androids and humanoids: closest to humans, and so with the most expectations to live up to.
• Zoomorphic: a familiar animal such as a cat, or an animal people know but have never handled, such as a dinosaur or a seal.
• Minimalist: robots such as Keepon and Muu that explore the minimum needed to trigger social interaction. The most extreme example is a busker robot that is nothing but a pair of moving sandals on a box and a sign reading "Naked Invisible Guy".
• Robjects: designs that start from an object rather than a creature, such as a social rubbish bin or a robotic ottoman.

Affordances and design principles

Norman's affordance is a visible relationship that tells you what actions are possible. A robot's appearance is a major affordance, because people assume the robot's abilities will match its looks. The textbook (section 4.1.4) distils the following principles.

• Match form to function. A humanoid is expected to talk, think and act like a human. For a job such as cleaning that needs none of this, a less humanlike design is better.
• Underpromise and overdeliver. Raising expectations through appearance or an introduction as an "intelligent companion", and then failing to deliver, causes deep disappointment and sometimes wrecks the interaction. Because the word "robot" itself makes the public expect high capability, the textbook suggests even considering not calling the artefact a robot.
• Interaction expands function. People fill in what the design leaves open with their own values and beliefs. Paro evokes a pet without being compared to any animal people know, so it is accepted as a pet-like creature even though its abilities fall far short of a real baby seal.
• Do not mix metaphors. A humanlike robot with skin covering only part of its body, or an animal robot that speaks like an adult and teaches mathematics, is odd. The Philip K. Dick android with its mechanics exposed is more unsettling because it is humanlike and machinelike at once (this connects to the uncanny valley in chapter 3).

Anthropomorphism

"My computer hates me." People attribute human traits, emotions and intentions to things that are not human. This is anthropomorphization. Seeing faces in random patterns, pareidolia, is one instance: the "face" on Mars photographed by Viking 1 in 1976 turned out to be plain terrain when re-photographed at higher resolution in 2001. HRI designers use this tendency as a design resource. Keepon expresses attention and affect with nothing more than two eyes and a symmetrical body.

In early HRI, anthropomorphism meant little more than how humanlike a robot looked. Epley et al. (2007) widened it into a psychological process with three factors: effectance motivation, the wish to explain the behaviour of an unfamiliar agent; sociality motivation, the wish for social connection; and elicited agent knowledge, the ease with which knowledge about humans comes to mind. People try to understand non-human things through the knowledge structures they have about themselves, and so attach feelings and intentions to robots. This changes not only impressions of the robot but actual behaviour towards it.

The media equation

Reeves and Nass (1996) showed in a series of HCI experiments that people treat computers as social actors. When a computer asked about its own performance, people rated it more generously than when another computer asked, which is to say they were polite to it. The response is almost automatic. This "computers are social actors" (CASA) approach led to studies of gender attribution to robots and of mental models of robots (textbook section 10.1).

Mind perception, warmth and competence

Gray et al. (2007) split mind perception into two dimensions: agency (planning, thinking, self-control) and experience (feeling emotions, having hopes). The two core dimensions of social judgement about people, warmth and competence, are used for robots as well. The Godspeed questionnaire (Bartneck et al. 2009) is a widely used HRI-specific instrument for anthropomorphism, available in many languages.

The uncanny valley

Mori (1970) proposed that affinity rises as robots become more humanlike, then plunges into a valley at the point where they are almost, but not quite, human. His example is the prosthetic hand. One that imitates wrinkles, veins and even fingerprints looks real at a glance, but shaking it reveals a boneless feel and coldness, and the result is eeriness. Movement amplifies both peak and valley: a switched-off industrial robot is just an oily machine, but the moment it moves like a hand some affinity appears, while a moving prosthetic hand becomes eerier. Mori suggested that stopping at a moderate degree of humanlikeness, like a Bunraku puppet, is the safe design.

Recent work points to mismatched features as a cause of the valley: when features differ in their degree of humanlikeness, such as very humanlike skin with slightly off facial muscle movement, eeriness grows (Kätsyri et al. 2015). The effect has been replicated with zoomorphic robots and with robots that mix gender cues. There is also a novelty component: eeriness fades with repeated interaction, short or long (textbook section 8.4).

Joggobot, a 2012 jogging-companion drone, flew just ahead of the runner. When its developers tested whether above, behind, in front or beside was best, runners who had the drone behind them felt chased and enjoyed the exercise less. They much preferred to be the one doing the chasing. Where a robot is placed relative to a person is a major design element in HRI.

Proxemics

The anthropologist Hall called the way people use space in relation to other people proxemics and divided it into four zones: intimate distance (a few centimetres to about 0.5 m) for close relationships and private talk, personal distance (about 0.5–1.2 m), social distance (1.2 to about 4 m) between acquaintances and colleagues, and public distance (about 4 m and beyond) for impersonal settings such as a speech. Intimate plus personal distance make up the personal space that people regard as their own. The distances vary with age and culture; Hall illustrated the difference between high-contact cultures such as South America and low-contact ones such as the United States with the North American barricading themselves behind a desk and chair. Arrangement follows rules too: people seated side by side behaved more cooperatively, people seated face to face more competitively, and people in conversation usually stand at an angle to each other.

Space in groups: F-formations

Kendon (1990) called an arrangement in which two or more people face each other and focus on a shared space an F-formation. It consists of the o-space in the middle, the p-space where the participants stand and the r-space beyond; two people form a face-to-face, L or side-by-side arrangement, larger groups a circle. At a cocktail party a tightly closed circle is hard to join, while an opened circle reads as an invitation. A robot can use this to judge which groups it may approach, and can change a group's F-formation by turning its body (Kuzuoka et al. 2010). One should not walk between two people in conversation even when there is room.

Social navigation for robots

Localisation (SLAM), finding people and avoiding collisions (dynamic window approach, path planning) are well studied in robotics. The problem is that an algorithm that weaves nicely between obstacles is socially inappropriate near people. A robot that treats people as obstacles swerves only inches before hitting someone coming the other way in a corridor. It avoids the collision, but to the person it looks aggressive. People give way much earlier, signalling with their bodies which way they will go, and never enter each other's personal space. The textbook summarises the research findings:

• Most people prefer to interact with a robot at personal or social distance (Walters et al. 2005). When passing in a corridor, the robot should keep enough distance not to enter the intimate zone (Pacchierotti et al. 2006).
• Approaching from the front is good for starting a conversation, but for handing over an object a frontal approach was less preferred and failed more often (Dautenhahn et al. 2006; Shi et al. 2013).
• When following someone, tracing the person's trajectory feels more natural than taking a shortcut (Gockley et al. 2007).
• When a person approaches the robot, it has to respond in time. A late response is awkward, and the person may not approach again (Kato et al. 2015).

A robot's trajectory is also a way of signalling intent: passing slowly a few metres from a visitor to show it is available, or using legible motion whose goal can be read when handing something over (chapter 5).

Even without words, people convey a great deal through gaze, facial expression, gesture and posture. Nonverbal signals are mostly automatic and sometimes unconscious, so they are considered more honest than speech: greeting someone you dislike politely while making less eye contact and avoiding touch. The same holds in HRI. A person may say nice things about a robot while their body is tense, and a robot whose gestures do not match the rhythm of its speech looks awkward. If a robot stares straight ahead and does not react to what we say, the interaction collapses. The textbook calls implementing nonverbal signals a precondition for smooth HRI.

Gaze

Gaze signals interest, understanding and attention, and manages turn-taking in conversation, for instance by moving the eyes from one person to another to choose who speaks next. Joint attention, two partners looking at the same thing together, is a scaffold for children learning language and the basis of coordination in collaboration. Robot gaze cues produce effects similar to those between people. In a conversation with several people, a robot can choose the next speaker with its gaze (Mutlu et al. 2009), can assign the roles of addressee, bystander and overhearer (Mutlu et al. 2012), and people who received eye contact while a robot told a story remembered it better (Mutlu et al. 2006). Making eye contact and then deliberately looking away makes a robot seem more thoughtful (Andrist et al. 2014).

Gesture

Gestures are grouped by role: deictic gestures that point at something, iconic gestures such as spreading the arms to show a big ball, symbolic gestures such as a wave that carry meaning on their own, and beat gestures that follow the rhythm of speech like conducting. In Salem et al. (2013), an ASIMO that gestured while speaking was rated more humanlike and likeable, and participants were more willing to interact with it again. Interestingly, gestures that did not match the speech raised the robot's ratings further but lowered task performance. Today, models trained on hundreds of hours of video of people talking generate gestures to accompany speech; the meaning does not always fit, but the movement itself is hard to tell from a human's (Yoon et al. 2022).

Mimicry, touch and channels only robots have

Unconscious mimicry signals liking, and when a partner copies your posture you are more inclined to accept their proposals, the "chameleon effect" (Chartrand & Bargh 1999). Mimicry that is too obvious or too late, however, looks manipulative and backfires. A robot that copied people's facial expressions and matched their head movements collected more donations (Wills et al. 2016). Besides body parts such as limbs, ears and tails, robots can signal through channels people do not have: sound, light and colour.

Signalling intent through motion

When handing over objects, people prefer robot motion whose goal and intent can be read (Koay et al. 2007). The most energy-efficient path may be unreadable to a person, so Dragan et al. (2013) proposed an algorithm that makes a robot arm produce legible motion towards its goal. They distinguish two concepts: predictable motion is the most efficient motion an observer who knows the goal would expect, while legible motion lets an observer infer the goal quickly while watching. The two sometimes conflict: to be more legible, the path has to be exaggerated away from the other candidate goals, which costs efficiency.

A robot in an electronics shop asks, "Hello, what are you looking for?" You answer, "Um, well, a camera for my daughter and some batteries, and I'm also just browsing." After a long silence the robot repeats the question and asks you to speak more slowly and closer to it. Why is talking to a robot so hard? Spoken interaction has several stages, and each has its own difficulties.

Speech recognition

Automatic speech recognition (ASR) turns recorded speech into text. It is transcription, not understanding. The technology assumes a quiet environment and a directional microphone, assumptions that often break in HRI: the person stands away from the robot, and ambient noise and the robot's own motors are recorded too. Lapel microphones and headsets are therefore common. Since deep learning, performance has risen sharply; in one test the textbook cites, the word error rate (WER) was 1.4%, below the human average of 5.8% (Zhang et al. 2022). Dialects, non-native speakers, children's speech, proper names and noisy, echoing rooms remain hard. Restricting what has to be recognised to numbers or short commands improves performance. ASR engines usually return a confidence score with the transcript, and sometimes alternative transcripts. In very noisy places, voice activity detection (VAD) alone can give the impression that the robot is listening and let it hand over the turn.

Extracting meaning and managing dialogue

Getting meaning out of the transcribed text is a separate problem. Sentiment analysis reduces an utterance to a single number for how positive it is; intent recognition spots requests such as "play some music". For tasks that take several turns, a dialogue manager (DM) tracks the state of the conversation; for a pizza order, the state is which pizza, size and toppings have been settled. In a real transcript the textbook quotes, a participant talking to a waiter robot that took everything literally had to rephrase several times, starting from "Could I have a glass of water?", before finding a wording the robot understood (Williams et al. 2018).

Turn-taking

While listening, people signal that they are following with back-channel signals such as "uh-huh" and "really?" and with nods. A robot can imitate this with short sounds, blinking lights and nods. Human responses are fast: in an analysis of telephone conversations, a "yes" came on average after 100 ms, and an answer to an unwelcome proposal after almost 500 ms (Strömbergsson et al. 2013). People prepare their answer before the question has ended. A robot chains recognition, understanding and generation and routinely takes seconds. Fillers and visual signals can bridge the silence but cannot replace fast turn-taking, so research looks at incremental synthesis that starts speaking before the sentence is planned, and incremental dialogue processing that starts acting before an instruction has finished.

Producing speech

Text-to-speech (TTS), turning text into sound, has advanced greatly over the last decade and is now hard to tell from a human voice. Traditional parametric synthesis is flexible but less natural; concatenative synthesis, which stitches recordings together, is natural but tied to the voice actor who was recorded. Generative models have moved past both limits. What remains is adapting to context: a robot speaks in the same voice in a quiet room and in a noisy exhibition hall.

Emotion is a universal communication channel that makes an inner state visible. Emotions differ from moods: an emotion has a clear cause and object and passes quickly, while a mood is more diffuse and its cause less clear.

How to represent emotion

Lists of emotions differ between people and cultures, so no single list is agreed on. Ekman and Friesen (1975) compiled six facial expressions recognised across cultures. This is often mistaken for a list of basic emotions, but it was originally a set of expressions that are shown on the face and recognised in many cultures. Russell (1980) argued that emotions are not discrete categories but interpretations of two independent neurophysiological systems, arousal and valence, and proposed the circumplex model, which spreads emotions over a two-dimensional plane. The model has been confirmed in many languages and cultures, although a meta-analysis notes that not every affective state lands where expected (Remington et al. 2000). Because the circumplex places "angry" and "afraid" side by side, the three-dimensional PAD model added a dominance axis; Kismet and many other robots use it. The OCC model defines 22 emotion categories through appraisal of situations, but many designers implement only Ekman's six expressions: reliably recognised, but narrow in interactional range.

Expression by robots

Ekman's Facial Action Coding System (FACS) groups facial muscles into action units (AUs) and writes an expression as a combination of AUs. A smile is AU6, which raises the cheeks, plus AU12, which pulls up the corners of the mouth. A robot face with rubber skin and 19 pneumatic actuators makes expressions from these AUs (Hashimoto et al. 2013), and many robots such as Kismet, iCat and eMuu use simplified facial cues. Non-humanlike robots express emotion too: a cleaning robot and a flying robot conveyed affect through movement patterns alone (Saerbeck & Bartneck 2010; Sharma et al. 2013), and speed, posture, sound, colour and orientation towards the person all serve as means. The facial feedback hypothesis holds that expressions in turn influence the emotion felt: people made to hold a pencil between their teeth, forcing a smile, found cartoons slightly funnier (Strack et al. 1988).

Why reading emotion is hard

Reading emotion from a still photograph of a face alone is nearly impossible even for people; whether a tennis player has just won or lost a point is hard to tell from the face by itself. Context, moving expressions and gestures raise recognition rates for people and algorithms alike. Most emotion-recognition software outputs only probabilities for the six basic expressions or a point in a two- or three-dimensional space, missing everyday emotions such as pride, embarrassment, guilt and irritation. The same emotion differs in intensity and context from person to person, so a robot has to be sensitive to the individual and the culture.

What trust is

Definitions of trust differ across psychology, sociology, economics and philosophy, but all contain the confidence that the other party will act appropriately. The textbook quotes Sabel's (1993) definition: "the mutual confidence that no party will exploit the other's vulnerability". It implies believing in the other's goodwill, integrity and ability (Lee & See 2004). In automation research, trust is what determines how much people rely on a system, and the key is not the amount of trust but its calibration. Trusting a system beyond its real ability is overtrust, which lets errors slip through; trusting it less is distrust, which throws away useful help. Hancock et al.'s (2011) meta-analysis grouped the factors behind trust in HRI into robot, human and environmental factors, and found the robot's performance-related factors the largest. The 2021 follow-up meta-analysis gives more weight to human factors. There are several scales for measuring trust, but no agreed definition yet.

Acceptance

A social robot has to be accepted by its users. Robot acceptance research mostly leans on the classic technology acceptance model (TAM): the intention to use depends on perceived usefulness and perceived ease of use. Because it treats the robot as a tool and leaves out context, extensions appeared: UTAUT adds social influence and facilitating conditions to performance and effort expectancy, and the Almere model (Heerink et al. 2010) adds affective factors such as trust, enjoyment and attitude for the acceptance of assistive technology by older adults. De Graaf et al. (2019) propose a user-experience (UX) perspective that includes hedonic factors, social norms and control beliefs.

Likeability and novelty

Affective cues and movement synchronised with the user raised a robot's likeability. But the response at a first encounter differs from the long-term response because of the novelty effect: people are unfamiliar with robots and are at first more enthusiastic or more anxious, and this settles over time. Short studies therefore tell us about first encounters, not about living with a robot (chapter 10).

Robots in closed environments, such as warehouse and delivery robots, are already commercial successes. The next wave is open environments with people in them: customer service, companionship, and social and physical assistance. Here an understanding of HRI decides success or failure (textbook ch. 11).

• Customer service: robots in malls and shops draw people in, at least while the novelty lasts. The museum tour guide was one of the first HRI applications, driving autonomously between exhibits and explaining them (Burgard et al. 1998); emotional expression such as a frustrated face when asking people to move helped the guiding. Hotel reception and in-store promotion are others.
• Education: the role of teacher's assistant was preferred by teachers and students alike (Reich-Stiebert & Eyssel 2016). Robots also appear as peers. One-to-one tutoring is highly effective (Bloom's 1984 two-standard-deviation claim has since been scaled back, but the benefit is clear), and the robot's physical presence was found to help learning more than an on-screen agent (Kennedy et al. 2015; Leyzberg et al. 2012). Why is not yet clear.
• Entertainment: companion and toy robots, exhibitions, performing arts.
• Health and therapy: long-term studies reported psychological, physiological and social effects of the seal robot Paro on older adults in care homes (Wada & Shibata 2007). In Japan individuals buy it; in Europe and the United States almost only institutions do. Children and adults with autism spectrum disorder generally respond well to robots, so there is much research on therapy support. Rehabilitation robots exist too.
• Service: cleaning robots, delivery robots in hospitals, hotels and restaurants. A restaurant robot must be designed not to block customers, or to let them step aside without irritation.
• Collaborative robots (cobots): traditional industrial robots were stiff, strong and senseless, so people could not go near them. Cobots work in the same space as people.
• Autonomous vehicles: signalling intent to pedestrians and other drivers is an HRI question.

Every application brings its own questions, such as the ethics of care robots (chapter 12).

The experimental approach has become the standard in HRI (Hoffman & Zhao 2020). Early HRI research used many methods that are embarrassing by today's standards, and there is a strong push to follow the standards of empirical sciences such as psychology (Baxter et al. 2016).

Research questions

A good research question replicates or extends prior work. Studies of human–human interaction are a blueprint for HRI; Reeves and Nass's CASA approach, which replicated classic results from psychology in HCI and then in HRI, is an example.

Kinds of methods

• Qualitative methods interpret qualities of an interaction that are hard to capture in numbers. They cannot establish correlation or causation, but they generate new hypotheses from rich data. The stopping criterion is not a fixed number of participants but saturation, the point where new data only repeat existing themes.
• Quantitative methods obtain numbers from questionnaires and controlled experiments to test correlation and causation.
• Mixed methods find factors through interviews and observation, then confirm them experimentally. If interviews suggest that a Roomba's autonomy makes it seem social, compare an autonomous Roomba with one driven by a gamepad.
• Others include user studies, surveys, system studies, observational studies, ethnography, conversation analysis, crowdsourcing and case studies.

Experimental design

What the researcher manipulates is the independent variable, and its values are its levels: a male and a female robot give two levels and two conditions. Factors that cannot be manipulated, such as participant gender, are quasi-experimental factors. What is measured is the dependent variable. A manipulation check is also needed to confirm that the manipulation worked, by asking whether participants actually perceived the robot's gender as intended. A between-subjects design assigns participants to conditions at random; in a within-subjects design the same people experience every condition, which reduces noise and needs fewer people. Some questions cannot use a within-subjects design, such as recovering from a broken leg with robot assistance, and there are order effects, such as always liking the first robot better, so the order of conditions has to be counterbalanced.

Measurement

Questionnaires commonly use Likert scales, which ask how much you agree with a statement, and semantic differential scales, which ask you to choose between opposites such as scary–friendly. Validated scales are better than home-made ones: the Godspeed questionnaire (Bartneck et al. 2009), which measures anthropomorphism, animacy, likeability, perceived intelligence and perceived safety, as well as RoSAS, the revised Godspeed and HRIES. Questionnaires are self-report, so they are complemented with implicit measures such as behavioural indicators, language use and physiological signals.

Wizard of Oz

When developing autonomy is not the point of the study, the Wizard of Oz (WoZ) technique has the research team control the robot from behind the scenes while it appears autonomous to participants. It suits technology that is almost usable but still unreliable, such as speech recognition. Participants must be told afterwards (debriefing), and faking unrealistic abilities, such as inventing an entire long conversation, is problematic. Riek (2012) proposed reporting guidelines for WoZ studies.

Participants, setting, duration

Sample size depends on the design and the target population and is calculated in advance with tools such as G*Power (chapter 11). Laboratory studies are easy to control; field studies capture the real context. Because first encounters differ from long-term interaction owing to the novelty effect, long-term studies spanning days to months, rarely years, are needed.

Descriptive statistics first

The tendency of the data (mean), the sample size and the variability (standard deviation) are the descriptive statistics. Always report the means, standard deviations (or ranges when the data are not normal), the number of participants, demographics, and any excluded data with the reason.

What a p-value says and does not say

Most classical tests combine tendency, sample size and variability into a single test statistic and produce a p-value: the probability of getting the present data if there were no real effect. The textbook's analogy is the probability of seeing this difference in ratings between two restaurants that are in fact equally good. This is the logic of null hypothesis significance testing (NHST). Two misunderstandings are common.

• A p-value is not the robustness of a result. If you replicate a significant result, the replication's p-value will, 80% of the time, land anywhere between 0.00008 and 0.44 (Cumming 2008). p-values are highly unstable across replications.
• A p-value is not the size or importance of an effect. Significance, size and importance are different things. An effect size (Cohen's d for t-tests, η² for ANOVA, adjusted R² for regression) measures how different two conditions are; the p-value adds the sample size to that. A rise in recovery rate from 4% to 6% is negligible for athlete's foot but important for survival of an aggressive cancer.

Reporting a 95% confidence interval conveys both the size of the effect and the precision of the estimate. The probability that a replication's mean falls inside the original study's confidence interval is about 83%.

Errors, power and p-hacking

Claiming an effect where there is none is a type I error; missing an effect that exists is a type II error. Both are reduced with sufficient power. Of the three things that determine power (tendency, variability, sample size), the researcher can only change the sample size. With p ≤ 0.05 as the criterion, one test in twenty comes out significant even without an effect, so running many tests eventually "finds" something. This is p-hacking. Psychology went through a replication crisis in which established effects failed to replicate, and HRI faces the same questions (Irfan et al. 2018). Widely used platforms such as Nao and Baxter make replication easier than before. Preregistration, registering a study in advance, prevents p-hacking and supports replication.

Robots in the media

People meet robots in stories before they meet them in person. Western narratives are full of robot uprisings (Čapek's R.U.R., the Terminator, the Matrix), and Asimov named the idea that robots will take over the world the "Frankenstein complex". Japan's Astro Boy and Doraemon are benevolent robots that help people, and this difference is often said to have played a part in Japan's development and acceptance of domestic robots. Either way, fictional robots colour people's expectations and responses. Recent research finds attitudes towards robots to be ambivalent rather than neutral or mildly positive (Stapels & Eyssel 2021, 2022): the hope that they will help sits alongside the fear of losing jobs or privacy.

Asimov's three laws

A robot may not injure a human being or, through inaction, allow one to come to harm; it must obey human orders unless they conflict with the first law; and it must protect its own existence as long as that does not conflict with the first two. A zeroth law protecting humanity as a whole was added later. The laws are widely known but have been criticised by philosophers, and Asimov's own stories show that they are not a practical guide.

Research ethics

The WoZ technique deceives participants, and can inflate expectations of robot ability so that people believe the technology is further along than it is (Riek 2012). Robots can also be used as persuasive technology that changes not just attitudes but behaviour, and the problem is that users are influenced without knowing how. Research with human participants needs approval from an ethics board (IRB), informed consent and debriefing.

Care, deception and dignity

Suppose you give your grandmother a companion robot. Sparrow and Sparrow (2006) argue that even if a robot provided perfect emotional and physical care, handing care over to a machine would be unethical. A relationship has meaning only between two parties who can exchange affect and concern, and however perfect the imitation of care, it must not replace the real thing. Sharkey and Sharkey (2012) ask how care robots might harm or support older adults' autonomy and independence. Alongside these sit the views of the robot as a tool, as a socially capable tool, and as a relational partner.

Further issues

• Privacy: a robot in the home moves around carrying cameras and microphones.
• Jobs: robots replace some occupations and create new ones. The task is to help those displaced find new work.
• Robot abuse: there is concern that mistreating humanlike robots could carry over to people; more research is needed.
• Expectations of relationships: a robot that complies with every demand may change what people expect of human relationships.
• Autonomous weapons: there is clear opposition, such as the campaign to stop killer robots.

The textbook's conclusion is that robots, people and their interaction are part of a larger society. Stories and values differ between cultures, and they decide how people see robots and how robots will change society.

Further reading

• The remaining textbook chapters: how robots work (ch. 3), the future (ch. 13)
• How robots learn to act, and learning from human preferences: the Robot Learning guide
• Controlling robots with brain signals: the Brain–Computer Interfaces guide

Each chapter of this guide matched to the chapters of Bartneck et al. (2024), 2nd edition, and to the papers it draws on. The textbook can be read chapter by chapter for free on the authors' official site.

Textbook

• Bartneck, C., Belpaeme, T., Eyssel, F., Kanda, T., Keijsers, M., & Šabanović, S. (2024). Human-Robot Interaction: An Introduction (2nd ed.). Cambridge University Press. free official edition from the authors

Papers

• Bartneck, C., Kulić, D., Croft, E., & Zoghbi, S. (2009). Measurement instruments for the anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots. International Journal of Social Robotics, 1(1), 71–81.
• Baxter, P., Kennedy, J., Senft, E., Lemaignan, S., & Belpaeme, T. (2016). From characterising three years of HRI to methodology and reporting recommendations. HRI 2016, 391–398.
• Chartrand, T. L., & Bargh, J. A. (1999). The chameleon effect: The perception–behavior link and social interaction. Journal of Personality and Social Psychology, 76(6), 893–910.
• Cumming, G. (2008). Replication and p intervals: p values predict the future only vaguely, but confidence intervals do much better. Perspectives on Psychological Science, 3(4), 286–300.
• Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319–340.
• Dragan, A. D., Lee, K. C. T., & Srinivasa, S. S. (2013). Legibility and predictability of robot motion. HRI 2013, 301–308.
• Ekman, P., & Friesen, W. V. (1978). Facial Action Coding System. Consulting Psychologists Press.
• Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886.
• Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619.
• Hall, E. T. (1966). The Hidden Dimension. Doubleday.
• Hancock, P. A., Billings, D. R., Schaefer, K. E., Chen, J. Y. C., de Visser, E. J., & Parasuraman, R. (2011). A meta-analysis of factors affecting trust in human-robot interaction. Human Factors, 53(5), 517–527.
• Hancock, P. A., Kessler, T. T., Kaplan, A. D., Brill, J. C., & Szalma, J. L. (2021). Evolving trust in robots: Specification through sequential and comparative meta-analyses. Human Factors, 63(7), 1196–1229.
• Heerink, M., Kröse, B., Evers, V., & Wielinga, B. (2010). Assessing acceptance of assistive social agent technology by older adults: The Almere model. International Journal of Social Robotics, 2(4), 361–375.
• Kätsyri, J., Förger, K., Mäkäräinen, M., & Takala, T. (2015). A review of empirical evidence on different uncanny valley hypotheses: Support for perceptual mismatch as one road to the valley of eeriness. Frontiers in Psychology, 6, 390.
• Kendon, A. (1990). Conducting Interaction: Patterns of Behavior in Focused Encounters. Cambridge University Press.
• Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.
• Mori, M. (1970/2012). The uncanny valley (K. F. MacDorman & N. Kageki, Trans.). IEEE Robotics & Automation Magazine, 19(2), 98–100. (original: Energy, 7(4), 33–35, 1970)
• Mutlu, B., Shiwa, T., Kanda, T., Ishiguro, H., & Hagita, N. (2009). Footing in human-robot conversations: How robots might shape participant roles using gaze cues. HRI 2009, 61–68.
• Reeves, B., & Nass, C. (1996). The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge University Press.
• Riek, L. D. (2012). Wizard of Oz studies in HRI: A systematic review and new reporting guidelines. Journal of Human-Robot Interaction, 1(1), 119–136.
• Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161–1178.
• Salem, M., Eyssel, F., Rohlfing, K., Kopp, S., & Joublin, F. (2013). To err is human(-like): Effects of robot gesture on perceived anthropomorphism and likability. International Journal of Social Robotics, 5(3), 313–323.
• Sharkey, A., & Sharkey, N. (2012). Granny and the robots: Ethical issues in robot care for the elderly. Ethics and Information Technology, 14(1), 27–40.
• Sparrow, R., & Sparrow, L. (2006). In the hands of machines? The future of aged care. Minds and Machines, 16(2), 141–161.