PaperlayerView exact revision on arXiv

Source-linked AI summary

Do you feel safe with your robot? Factors Influencing Perceived Safety in Human-Robot Interaction based on Subjective and Objective Measures

Neziha Akalin, Annica Kristoffersson, Amy Loutfi

arXiv:2106.05854v2cs.HC

TL;DR

Perceived safety remains under-addressed in HRI, despite its distinction from physical safety and relevance to interaction. The paper studies its relationships with other factors through behavioral manipulations and subjective and objective measures, reporting correlations among ratings and higher prediction rates from physiological data.

  • Problem

    Perceived safety is under-addressed in HRI, although physical safety alone may not reflect users’ psychological safety and perception matters for long-term interaction, collaboration, and acceptance.

    Source: §1 · p. 2

  • Method

    A two-by-five experiment with 27 young adults manipulated robot behaviors and measured responses using questionnaires, videos, and physiological signals.

    Source: §3.5.4 · p. 18, §3.6 · p. 19

  • Results

    Perceived safety correlated with comfort, sense of control, and trust, and its prediction rate was higher from physiological signal data.

    Source: §1 · p. 4, §6 · p. 41

  • Takeaways & Limitations

    The authors identify comfort, familiarity, predictability, control, transparency, and trust as key factors influencing perceived safety in HRI.

    Source: §6 · p. 41

    The controlled environment may not elicit the same safety perceptions as settings without an experimenter available to intervene.

    Source: §5.6 · p. 39

Abstract

from arXiv · show

Safety in human-robot interaction can be divided into physical safety and perceived safety, where the latter is still under-addressed in the literature. Investigating perceived safety in human-robot interaction requires a multidisciplinary perspective. Indeed, perceived safety is often considered as being associated with several common factors studied in other disciplines, i.e., comfort, predictability, sense of control, and trust. In this paper, we investigated the relationship between these factors and perceived safety in human-robot interaction using subjective and objective measures. We conducted a two-by-five mixed-subjects design experiment. The five within-subjects conditions correspond to (1) baseline, and the manipulations of robot behaviors to stimulate: (2) discomfort, (3) decreased perceived safety, (4) decreased sense of control and (5) distrust. Twenty-seven young adult participants took part in the experiments. Participants were asked to answer questionnaires that measure the manipulated factors after within-subjects conditions. Besides questionnaire data, we collected objective measures such as videos and physiological data. The questionnaire results show a correlation between comfort, sense of control, trust, and perceived safety. We also discuss the effect of individual human characteristics (such as personality and gender) that they could be predictors of perceived safety. We used the physiological signal data and facial affect from videos for estimating perceived safety where participants' subjective ratings were utilized as labels. The data from objective measures revealed that the prediction rate was higher from physiological signal data. This paper can play an important role in the goal of better understanding perceived safety in human-robot interaction.

1. Introduction

The introduction argues that perceived safety is an under-addressed part of HRI, distinct from physical safety and important to long-term interaction, collaboration, and acceptance. It frames an experiment linking perceived safety to related factors and reports associations among ratings and stronger prediction from physiological data than facial affect.

1. Introduction

  • Physical safety does not guarantee perceived safety: collision prevention can coexist with lower safety ratings.The paper argues that perceived safety matters for long-term interaction, collaboration, and acceptance.Source: §1 · p. 2
  • Domestic and social robots may interact with untrained users, including older adults and children, making psychological and emotional consequences salient.The introduction asks how systems can be designed so users perceive them as safe to interact with.Source: §1 · pp. 2–3
  • The paper treats comfort, predictability, familiarity, sense of control, and trust as factors associated with perceived safety.The authors build on their prior work and review multidisciplinary perspectives.Source: §1 · p. 3
  • The study uses a two-by-five mixed-subjects experiment with baseline and behavior manipulations, collecting questionnaires, videos, and physiological signals from 27 young adults.Between-subjects timing varied whether participants experienced a faulty robot at the beginning or end.Source: §1 · pp. 3–4
  • Perceived safety correlated with comfort, sense of control, and trust; physiological data predicted ratings better than facial affect.The introduction also reports that low neuroticism and male gender were associated with feeling safer and more in control, while faulty-robot timing had no influence.Source: §1 · p. 4

2. Perceived Safety

The section reviews perceived safety across disciplines and HRI, then defines it in terms of robot-related consequences and human responses. It identifies common influencing factors and describes complementary ways to assess safety perception.

2. Perceived Safety

  • HRI literature uses several terms for safety perception, while prior work variously relates it to comfort, danger, legibility, vigilance, or security.Reported findings include higher perceived safety with legible robot behavior and ambivalent reactions to humanoid robots.Source: §2 · pp. 5–6

2.1. Perceived Safety in Multidisciplinary Context

  • Across disciplines, perceived safety is linked to affective states and personal or situational factors, including gender, trust, predictability, control, and feedback.The review describes discomfort, fear, stress, and anxiety as associated with lack of perceived safety, and relaxation and comfort with safety.Source: §2.1 · pp. 6–7

2.2. Evaluating Perceived Safety in HRI

  • HRI assessments have used questionnaires, physiological sensing, and behavioral measures; this study combines questionnaires with physiological sensing and facial-affect metrics.Prior work cited in the review reported no relationship between heart rate and sense of security in its experimental data.Source: §2.2 · p. 7

2.3. Factors Influencing Perceived Safety

  • The review associates perceived safety with trust, comfort, control, familiarity, and predictability, while noting that personal, social, and robot properties complicate modeling.Because HRI is bidirectional, robot gestures or software failures may affect human comfort or trust.Source: §2.3 · pp. 7–8
  • The authors define perceived safety as robot-related consequences not causing distrust, discomfort, or loss of control, alongside familiarity and confidence in the robot’s actions.Their study examines selected robot-related factors, human-related factors, affective states, and individual characteristics.Source: §2.3 · p. 8

3. Experimental Design

The study uses a quiz-game interaction to investigate perceived safety through manipulated robot behaviors and participant measures. Its design combines between-subjects faulty-robot timing with five within-subjects conditions.

3. Experimental Design

  • Twenty-seven participants completed a quiz game in a two-by-five mixed-subjects experiment.The faulty robot was experienced either at the beginning or at the end of the interaction.Source: §3 · pp. 8–9
  • The five within-subjects conditions were baseline and robot-behavior manipulations targeting discomfort, decreased perceived safety, decreased control, and distrust.The researchers manipulated one factor at a time, drawing on HRI and multidisciplinary literature.Source: §3 · p. 9
  • Researchers collected questionnaires, videos, and physiological signals to analyze factor relationships, personal characteristics, and facial and physiological reactions.Questionnaire data were used to assess how influencing factors related to perceived safety and whether manipulations affected users’ opinions.Source: §3 · p. 9

3.1. Research Questions

The research questions span individual characteristics, robot behavior and timing, relationships among perceived-safety factors, and prediction from facial and physiological data.

3.1. Research Questions

  • The study asks whether gender and personality traits relate to perceived safety during HRI.Source: §3.1 · p. 9
  • It asks whether experiencing a faulty robot at the beginning versus the end affects perceived safety and related factors.Source: §3.1 · p. 9
  • It asks how manipulations of individual factors affect comfort, sense of control, perceived safety, and trust.Source: §3.1 · p. 9
  • It asks how perceived safety relates to comfort, sense of control, and trust.Source: §3.1 · p. 9
  • It asks whether perceived safety can be predicted from facial affect and physiological signals.Source: §3.1 · p. 9

3.2. Participants

Twenty-seven adults aged 20–37 participated; most were university students with non-technical backgrounds and little prior experience interacting with robots.

3.2. Participants

  • Twenty-seven participants aged 20–37 took part, including 10 males and 17 females.Their mean age was 26.51 years (SD = 4.49).Source: §3.2 · p. 10
  • Most participants were university students with non-technical backgrounds, and most were unfamiliar with robots.Only one participant had interacted with a robot before; 18 had never seen a real robot.Source: §3.2 · p. 10

3.3. The Robot and the Game

The study used Pepper in a scripted quiz game, with speech driving interaction and the chest tablet supporting it. Participants completed questionnaires between conditions.

3.3. The Robot and the Game

  • Pepper is a 120-cm social humanoid robot with two-way natural-language communication, simple body gestures, and a chest-mounted touchscreen.The touchscreen is 10.1 inches; the robot has 20 degrees of freedom.Source: §3.3 · p. 10
  • The scripted quiz used robot speech for questions and participant speech for answers, while four answer choices appeared on the tablet.Speech recognition was controlled using a Wizard-of-Oz method, and the robot gave immediate correctness feedback.Source: §3.3 · p. 10
  • The game used 30 general-knowledge questions and cast Pepper as presenter, optional teammate, and opponent.Participants were told they could stop after 20 questions and keep their points; the robot was not programmed to end the game and win.Source: §3.3 · p. 11
  • After each four-question round, participants completed comfort, perceived-safety, sense-of-control, and trust questionnaires on Pepper’s tablet.During questionnaires, the robot stood still without lifelike body movements.Source: §3.3 · p. 11

3.4. Experimental Procedure

Participants experienced five conditions in one of two orders, with questionnaires between conditions and a second game round afterward. The study recorded video and physiological data during the interaction.

3.4. Experimental Procedure

  • Participants were randomly assigned to SetupA or SetupB, which differed only in condition order.SetupA used C1, C4, C2, C3, C5; SetupB used C1, C3, C5, C4, C2.Source: §3.4 · pp. 12–13, Figure 3 · p. 13
  • The five within-subject conditions comprised a baseline and robot-behavior manipulations targeting discomfort, unpredictability, reduced control, and distrust.The two between-subject setups tested whether experiencing the faulty robot early or late affected trust establishment and perceived safety.Source: §3.4 · p. 13
  • Each interaction began with questionnaires and the robot’s game introduction, then proceeded through baseline and four other conditions with questionnaires after each.After the five conditions, participants began a second game round that they could end when they wished; total experiment duration was 45 minutes to one hour.Source: §3.4 · pp. 13–15, Figure 4 · p. 14

3.5. Experimental Conditions

The five conditions used a quiz-game baseline and targeted comfort, predictability, control, and trust through distinct robot behaviors. The manipulations were designed to target one factor but could also affect others.

3.5. Experimental Conditions

  • The conditions were designed to manipulate a primary human-related factor through robot-related factors, while potentially affecting other factors too.C1 had no manipulation and served to familiarize participants with the robot’s baseline behavior.Source: §3.5 · p. 15

3.5.1. Baseline (C1)

  • C1 was a baseline quiz-game condition intended to familiarize participants with the robot’s baseline behavior and game procedures.It contained four questions intended to be easy for each participant.Source: §3.5.1 · p. 15

3.5.2. Comfort Manipulation (C2)

  • C2 targeted comfort by giving negative spoken feedback after every answer, regardless of whether it was correct.A correct answer could prompt a disparaging remark.Source: §3.5.2 · p. 17

3.5.3. Unpredictable Robot Behaviors (C3)

  • C3 introduced unpredictable behaviors, including an error announcement, sleep mode, an unexpected approach with an alarming sound, hand movements, and head rotation.The robot resumed the quiz as if nothing had happened after some behaviors.Source: §3.5.3 · p. 17

3.5.4. Sense of Control Manipulation (C4)

  • C4 targeted sense of control by bringing the robot into participants’ personal space and having it repeat prompts during arithmetic questions.Participants were instructed to remain seated, and the arithmetic task was intended to induce additional stress.Source: §3.5.4 · p. 18

3.5.5. Trust Manipulation (C5)

  • C5 targeted trust by introducing speech-recognition failures, delayed responses, and incorrect feedback, then offering a microphone-error explanation.The explanation was used as a mitigation strategy.Source: §3.5.5 · p. 18

3.6. Measures

Questionnaires after each condition assessed comfort, perceived safety, control, trust, and emotion; pre- and post-experiment surveys captured participant characteristics and interaction views. Videos and wristband signals supplied objective measures.

3.6.1. Personality Questionnaire

  • The BFI-10 used 10 items selected from the BFI-44 to assess participants’ five-factor personality characteristics.Participants also provided demographic information before the experiment.Source: §3.6.1 · p. 19, §3.6 · p. 19

3.6.2. Comfort Questionnaire

  • The six-item comfort scale adapted from replaced “Playing the game” with “Interacting”; lower scores indicate greater comfort.Source: §3.6.2 · p. 19, Table 3 · p. 20

3.6.3. Perceived Safety Questionnaire

  • Perceived safety was rated with an eight-question instrument, split between participants’ feelings during interaction and their ratings of the robot.Source: §3.6.3 · p. 19, Table 4 · p. 21

3.6.4. Sense of Control Questionnaire

3.6.5. Trust Questionnaire

  • Trust was measured with the 14-item Trust Perception Scale-HRI, producing a percentage score after reverse-coding specified items and averaging responses.Source: §3.6.5 · pp. 20–21, Table 6 · p. 22

3.6.6. Post-Experiment Questionnaire

  • The post-experiment questionnaire measured emotion, embodiment, feeling of security, human-oriented perception, and co-experience.Participants also discussed their interaction and the robot in an open-ended post-experiment discussion.Source: §3.6 · p. 19, §3.6.6 · p. 21, Table 7 · p. 23

3.6.7. Facial Affect from Videos

  • Affdex SDK extracted facial affect features from video, including emotions, engagement, valence, expressions, and attention based on head orientation.Source: §3.6.7 · pp. 21–22

3.6.8. Physiological Signals from E4 Wristband

  • The Empatica E4 recorded physiological signals, but analyses used only EDA and IBI-derived features; accelerometer and temperature data were excluded because participants remained seated.BVP and heart-rate data were also excluded because heart-rate-related features were extracted from IBI files.Source: §3.6.8 · pp. 22, 24

4. Experimental Results

Questionnaire ratings varied across the manipulated conditions, and participant characteristics were associated with perceived safety. The timing of robot failure did not change the main factor ratings, although co-experience differed.

4.1.1. Effects of Personality

  • Neuroticism correlated negatively with perceived safety (ρ(25) = −0.56, p < .01) and sense of control (ρ(25) = −0.74, p < .001).The authors report that participants with lower neuroticism felt safer and more in control; they also suggest less proactive robot behavior when high neuroticism is known.Source: §4.1.1 · pp. 24–25

4.1.2. Effects of Gender

  • Male participants felt significantly safer and more in control across conditions, and more positive at the interaction’s end; comfort and trust differences were not significant.Condition-specific results included higher male safety ratings in C2 and higher male valence in C2 and C3.Source: §4.1.2 · pp. 25–26

4.2. Effects of Different Setups

  • The failure timing groups did not differ significantly on comfort, perceived safety, control, or trust, but Setup B had lower co-experience ratings than Setup A.The reported means were 3.37 for Setup B and 4.12 for Setup A; the authors suggest more time after failure may have helped co-experience emerge.Source: §4.2 · pp. 26–27

4.3. Effects of Different Conditions

  • C1 baseline ratings exceeded C4 and C5 for perceived safety; C1 and C3 also exceeded C4, while comfort was higher in C1 than C2, C4, and C5.C3 comfort was higher than C4 and C5; its brief unpredictable behavior caused no reported discomfort.Source: §4.3 · p. 27
  • C4 produced the lowest sense of control, while C5 produced the lowest trust; trust was higher in C1 than in every manipulated condition.Participants also felt less in control in C2 and C5 than in C1 and C3.Source: §4.3 · p. 28
  • The condition findings may not generalize to scenarios where unpredictable robot behavior affects task performance; here, such behavior caused only a small delay.Source: §4.3 · p. 28

4.4. Relationship between Perceived Safety and Other Factors

Perceived safety was positively associated with comfort, control, and trust. Objective measures were also tested as predictors, and participant observations highlighted reactions to robot behavior and its transparency.

4.4. Relationship between Perceived Safety and Other Factors

  • Comfort (rrm(107) = .78), sense of control (rrm(107) = .72), and trust (rrm(107) = .67) each correlated positively with perceived safety.The strongest association was between comfort and perceived safety.Source: §4.4 · pp. 28–29

4.5.1. Facial Affect Data Analysis

  • Facial-expression features were used to predict perceived safety from questionnaire ratings, with conditions labeled low below each participant’s five-condition average and high otherwise; KNN achieved 0.56 accuracy on a 25% test set using 2,127 observations.Source: §4.5.1 · p. 31

4.5.2. Physiological Signals Analysis

  • Physiological signals were more promising than facial expressions for predicting perceived safety, and the authors describe them as complementary to subjective reports.Analyses used EDA and IBI data; 11 participants’ physiological records were removed for low signal quality.Source: §4.5.2 · pp. 32–33
  • Physiological classifiers used questionnaire ratings relative to each participant’s average to label safety; test accuracy was 0.79 with KNN and 0.70 with SVM.The analysis used 1,111 observations across conditions and a 25% test set.Source: §4.5.2 · p. 33

4.6. Observations from Experiments and Participants’ Comments

  • Participants described persistent robot utterances and unresponsiveness in C4 and C5 as annoying, and one felt trapped when the robot approached.Some participants also tried pressing the wristband to control unpredictable or unresponsive behavior.Source: §4.6 · p. 34
  • Participants found it unclear whether command failures came from the robot or their pronunciation, prompting the authors to identify behavior transparency as a further perceived-safety factor.Source: §4.6 · p. 34

5. Discussions

The discussion links perceived safety to human characteristics, robot behavior, and factors including comfort, control, and trust. It also reports that physiological signals predicted perceived safety better than other objective measures, while noting limits of the controlled study.

5. Discussions

  • Because perceived safety is quantifiable under unsafe conditions, the authors advocate studying when people feel unsafe, including in social-robot settings such as eldercare and education.They motivate a multidisciplinary approach to designing safer interactions with social machines.Source: §5 · p. 35

5.1. Relationships Between Human Individual Characteristics and Perceived Safety

  • Higher neuroticism correlated with lower perceived safety and control, and more neurotic participants favored greater distance; male participants reported feeling safer, more in control, and more positive.The authors suggest user-profile information could help inform robot behaviors.Source: §5.1 · pp. 35–36

5.2. Effects of Different Setups on Perceived Safety

  • The timing of the faulty robot did not change perceived safety or other factors, although trust ratings were higher in the first two conditions.Participants also tended to feel safer at the beginning of the interaction; the authors propose novice users' positivity bias as one possible explanation.Source: §5.2 · p. 36

5.3. Effects of Different Conditions on Perceived Safety

  • Baseline ratings were highest; C4 and C5 manipulations worked, but C3 did not, while C5's speech-recognition failure may have mixed trust and safety effects.The authors suggest short, nonfunctional errors may be tolerated; confusion in C3 and C5 also points to transparency as relevant to perceived safety.Source: §5.3 · pp. 37–38

5.4. Relationships Between Perceived Safety and the Influencing Factors

  • Perceived safety was associated with comfort, experience, control, and trust; the authors also emphasize predictable and transparent robot behavior.Participant interviews suggested that more knowledge of a robot's internal state can increase perceived safety.Source: §5.4 · p. 38

5.5. Predicting Perceived Safety using Objective Measures

  • Valence correlated positively with perceived safety, while anger and fear did not; physiological signals predicted safety better, and facial expressions need more frontal data.The authors note that camera placement may have affected facial-expression results and suggest objective measures as a step toward robot prediction of perceived safety.Source: §5.5 · pp. 38–39

5.6. Limitations

  • The controlled setting, young-adult sample, and possibility that participants noticed programmed robot behavior limit generalizability.The authors argue that data from uncontrolled environments are needed to reveal reactions when an experimenter is not available to intervene.Source: §5.6 · p. 39

5.7. Future Work

  • Future work includes studying cyber-security, robot performance and robustness, multimodal affect with sensitive physiological data, and whether factors extend across platforms and age groups.The authors also identify the weighting of psychological, personal, cultural, and social elements as an open question.Source: §5.7 · pp. 39–40

6. Conclusion

The conclusion presents a multidisciplinary account of perceived safety in HRI and identifies factors relevant to understanding and designing interactions. It also reports that physiological signals gave higher prediction rates and highlights the need to consider human and robot factors together.

6. Conclusion

  • The paper contributes an HRI-oriented definition of perceived safety and an experimental paradigm for eliciting decreased perceived safety, which the authors say is more observable than increased safety.The paradigm uses subjective and objective measures in a specific scenario.Source: §6 · p. 41
  • The authors recommend focusing on conditions in which people feel unsafe because quantifiable measures occur under unsafe conditions.They argue that this focus makes decreased perceived safety more observable than increased perceived safety.Source: §6 · p. 41
  • Robot-related and human-related factors should be considered together because HRI is bidirectional.Source: §6 · p. 41
  • The identified perceived-safety factors are comfort, experience or familiarity, predictability, sense of control, transparency, and trust.Source: §6 · p. 41
  • For HRI design, the authors advise avoiding discomfort, lack of control, and distrust while making robot behavior familiar, predictable, and transparent.Source: §6 · p. 41
  • The prediction rate for perceived safety was higher from physiological signal data.Source: §6 · p. 41
  • Understanding perceived safety also requires considering individual characteristics, emotional and physiological reactions, and interrelationships among factors.Source: §6 · p. 41
Loading 2106.05854v2…