Source-linked AI summary
Expectations, Explanations, and Embodiment: Attempts at Robot Failure Recovery
Elmira Yadollahi, Fethiye Irmak Dogan, Yujing Zhang, Beatriz Nogueira, Tiago Guerreiro, Shelly Levy Tzedek, Iolanda Leite
TL;DR
The paper examines how robot explanations are perceived under different user expectations, a relationship that remains underexplored in HRI. Across two stages, it validates video priming and evaluates explanations after robot failures; explanations improved perceptions especially for Furhat and under low expectations.
Problem
How robot explanations are perceived under different user expectations, including expectations shaped by priming, remains underexplored in HRI.
Source: §1 · p. 3, §2.4 · p. 6Method
Results
Explanations improved robot perceptions, particularly for Furhat and among negatively primed participants, where they more strongly enhanced expressiveness and explanation satisfaction.
Source: §1 · p. 4, §6 · p. 23Takeaways & Limitations
Explanation effects varied by robot, and explanations were especially helpful when participants had lower expectations.
Source: §6 · p. 22, §8 · p. 26The online video-based evaluation leaves ecological validity and generalization to in-person or prolonged interactions uncertain.
Source: §6.1 · pp. 24–25
Abstract
from arXiv · showhide
Expectations critically shape how people form judgments about robots, influencing whether they view failures as minor technical glitches or deal-breaking flaws. This work explores how high and low expectations, induced through brief video priming, affect user perceptions of robot failures and the utility of explanations in HRI. We conducted two online studies ($N=600$ total participants); each replicated two robots with different embodiments, Furhat and Pepper. In our first study, grounded in expectation theory, participants were divided into two groups, one primed with positive and the other with negative expectations regarding the robot's performance, establishing distinct expectation frameworks. This validation study aimed to verify whether the videos could reliably establish low and high-expectation profiles. In the second study, participants were primed using the validated videos and then viewed a new scenario in which the robot failed at a task. Half viewed a version where the robot explained its failure, while the other half received no explanation. We found that explanations significantly improved user perceptions of Furhat, especially when participants were primed to have lower expectations. Explanations boosted satisfaction and enhanced the robot's perceived expressiveness, indicating that effectively communicating the cause of errors can help repair user trust. By contrast, Pepper's explanations produced minimal impact on user attitudes, suggesting that a robot's embodiment and style of interaction could determine whether explanations can successfully offset negative impressions. Together, these findings underscore the need to consider users' expectations when tailoring explanation strategies in HRI. When expectations are initially low, a cogent explanation can make the difference between dismissing a failure and appreciating the robot's transparency and effort to communicate.
1. Introduction
Robot failures can disappoint users when expectations exceed robots’ capabilities, and the role of explanations in this mismatch remains underexplored. The paper tests whether brief video priming shapes expectations and whether explanations affect perceptions of failures.
1. Introduction
- Attribution Theory and Expectancy-Disconfirmation Theory suggest that explanations of a robot’s reasoning or constraints may mitigate disappointment after failure or underperformance.High expectations followed by failure can produce negative disconfirmation and reduce trust and satisfaction.Source: §1 · pp. 2–3
- The study addresses limited evidence on how user expectations affect perceptions of robot explanations for capability mismatches.Prior work examined expectations and explanations separately, without accounting for expectations during explanation generation.Source: §1 · p. 3
- The research used two studies replicated with Pepper and Furhat, first validating brief videos intended to induce high or low expectations.The second study compared failure scenarios with and without robot explanations.Source: §1 · pp. 3–4
- Explanations generally improved robot perceptions, especially for Furhat; after low-expectation priming, they more strongly improved perceived expressiveness and explanation satisfaction.The priming videos successfully induced the intended expectations.Source: §1 · p. 4
2. Related Work
The related work frames expectations as shaping robot evaluations, reviews priming and failure research, and identifies limited evidence on how expectations shaped by priming affect perceptions of robot explanations.
2.1. Role of Expectations
- Expectancy-Disconfirmation Theory describes satisfaction as depending on whether performance meets or diverges from expectations, while expectation-capability mismatches can bring disappointment.Expectations are described as guiding behaviour, hopes, and intentions.Source: §2.1 · pp. 4–5
2.2. Role of Priming
- Priming is presented as a way to shape expectations and behaviour, while prior HRI work examined emotional priming and robot portrayals in media.Liao and MacDonald (2020) examined emotional priming in autonomous products; Banks (2020) studied sympathetic robot portrayals.Source: §2.2 · p. 5
2.3. Role of Robot Failures
- Robot failures can be technical or interactional, and prior studies examined responses to conversational failures and explanations for plan-execution failures.Attribution Theory offers a framework for how users interpret failures, including assigning causes to internal or external factors.Source: §2.3 · pp. 5–6
2.4. Role of Explanations
- Explanations can support transparency and understanding, but unclear explanations may harm interaction; how expectations shaped by priming affect explanation perceptions remains underexplored.The cited work distinguishes what-, why-, and how-explanations, and notes that clear why-explanations are often needed when robots behave unexpectedly.Source: §2.4 · p. 6
3. Priming Study: Validating Priming Effect
The priming study was designed to test whether short interaction videos could shift perceptions of Pepper and Furhat, whose embodiments and expressive capabilities differ. Positive and negative scenarios were intended to establish contrasting expectations for later evaluations.
3. Priming Study: Validating Priming Effect
- The study tested whether brief interaction videos could shift perceptions of two contrasting robots: mobile Pepper, with a static face, and stationary Furhat, with a nuanced expressive face.The robots were selected to examine whether embodiment differences affected priming and subsequent evaluations.Source: §3 · p. 7
3.1. Study Objective
- The objective was to shift robot-capability perceptions positively or negatively and assess priming through shared-perception, interpretation, and nonverbal-expressiveness scales.Positive videos showed flawless task execution and social interaction, while negative videos showed failures intended to reduce confidence in competence.Source: §3.1 · p. 8
3.2. Design of Priming Videos
- Restaurant scenarios tailored to each robot depicted communication, hardware, or social-norm failures in negative clips and matched successful interactions in positive clips.The scenarios were kept as comparable as possible; each video lasted about two minutes and included subtitles.Source: §3.2 · pp. 8–10
3.3. Priming Study Design
The priming study used positive and negative videos for Furhat and Pepper, with questionnaires measuring perceived robot capabilities and general attitudes. Its design also included a Prolific sample and distinct scales for embodiment-related perceptions and robot attitudes.
3.3. Priming Study Design
- Participants were randomly assigned to positive or negative priming videos for Furhat or Pepper in a between-subjects design.The study recruited 208 Prolific participants against a power-analysis target of 144; eight who failed an attention check were excluded.Source: §3.3 · p. 10, §3.4 · p. 10
3.4. Participants and Power Analysis
- The study recruited 208 Prolific participants against a power-analysis target of 144, and excluded eight who failed the attention check.The reported sample included participants aged 18–67 and varied levels of robot familiarity.Source: §3.4 · pp. 10–11
3.5.1. The Peculiarities of Robot Embodiment (EmCorp-Scale)
- EmCorp measures perceived interpretation, mobility, and nonverbal expressiveness, including perceived vision and hearing, movement and object manipulation, and expressive cues.Interpretation, Mobility, and Expressiveness are the study's shortened labels for these constructs.Source: §3.5.1 · p. 11
3.5.2. General Attitudes Towards Robots Scale (GAToRS)
- GAToRS measures positive and negative attitudes at personal and societal levels with 20 items rated on a 7-point scale.Its subscales cover personal trust and unease, and perceived societal benefits and concerns.Source: §3.5.2 · pp. 11–12
3.6. Study Procedure
Participants completed baseline attitude questions, watched a priming video, rated the robot, and then repeated the attitude questions. An attention check was embedded in the perception questionnaire.
3.6. Study Procedure
- Participants completed pre-GAToRS, watched a positive or negative priming video, rated the robot with EmCorp, and completed post-GAToRS.The pre/post attitude measures were used to assess change following the video.Source: §3.6 · p. 12
- The procedure included demographic and robot-familiarity questions before priming and an attention check embedded in the EmCorp questionnaire.The attention-check item expected disagreement with a statement participants had been prompted to agree with.Source: §3.6 · p. 12
3.7. Priming Results
Positive priming increased perceived interpretation and expressiveness for both robots. General attitudes also changed, with different significant subscales across robot and priming conditions; mobility was excluded from analysis because it did not suit Furhat.
3.7.1. Data Preparation
- Mobility met acceptable internal-consistency levels but was excluded for both robots because Furhat lacks mobility.The reported alpha thresholds were at least 0.8 for interpretation and mobility and at least 0.6 for expressiveness.Source: §3.7.1 · p. 13
3.7.2. Effect of Priming on Robot Perception
- Positive priming significantly increased perceived interpretation and expressiveness for both Furhat and Pepper.For Furhat, interpretation medians were 4.0 versus 2.85 and expressiveness medians were 4.0 versus 3.75; Pepper's corresponding medians were 3.71 versus 2.71 and 4.0 versus 2.6.Source: §3.7.2 · pp. 13–14
3.7.3. Effect of Priming on General Attitudes
- For negatively primed Furhat, personal negative attitudes increased and personal positive attitudes decreased from pre- to post-priming.The reported medians changed from 3.2 to 3.6 for personal negative and from 4.8 to 4.5 for personal positive.Source: §3.7.3 · p. 14
- For positively primed Furhat, personal negative, societal positive, and societal negative attitudes changed significantly.The reported pre/post medians were 4.6/4.8, 5.4/5.4, and 5.0/5.2, respectively.Source: §3.7.3 · p. 14
- For Pepper, negative priming significantly changed personal negative attitudes, while positive priming changed societal positive attitudes.The positive-priming societal-positive median changed from 5.2 to 5.6.Source: §3.7.3 · pp. 14–15
3.8. Validating Priming Effect: Discussion
Positive priming raised perceived capability ratings for both robots, and priming also shifted broader robot attitudes. The discussion notes that these attitude effects were measured immediately, leaving their longevity uncertain.
3.8. Validating Priming Effect: Discussion
- Positive priming raised perceived interpretation and expressiveness for both robots, even though their behavior remained the same.The authors interpret this pattern as consistent with higher expectations enhancing perceived competence.Source: §3.8 · p. 15
- Priming also shifted broader attitudes, but their persistence is uncertain because participants completed questionnaires immediately after the videos.The discussion proposes investigating whether repeated or prolonged interactions sustain or erode these altered expectations.Source: §3.8 · p. 15
4. Main Evaluation: Effect of Failures and Explanation
The main evaluation tested whether primed expectations shape perceptions of robot failures and whether brief verbal explanations affect those perceptions. Participants saw a priming video followed by a failure scenario featuring the same robot, with or without explanations.
4. Main Evaluation: Effect of Failures and Explanation
- The evaluation examined how primed expectations and explanations affected robot perceptions after failures, while noting that effects beyond a short-term interaction were not measured.The study did not collect longitudinal data, leaving the persistence of effects open.Source: §4 · p. 15
4.1. Hypotheses
- The hypotheses predicted improved EmCorp perceptions after explanations and greater explanation satisfaction among negatively primed participants.Prior work cited by the authors links explanations with perceptions, competence, trust, and satisfaction.Source: §4.1 · p. 16
4.2. Design of Failure Videos with and without Explanations
- Main videos used new museum scenarios and actors; Pepper dropped an object, while Furhat’s errors centered on communication lapses.Videos included communication failures, hardware malfunctions, and social norm violations; explanations briefly apologized and stated the cause of each failure.Source: §4.2 · pp. 16–17
4.3. Study Design
- The between-subjects design crossed positive or negative priming with explanation or no explanation and was replicated with Furhat and Pepper.Participants watched a restaurant priming video followed by a museum failure video featuring the same robot; Figure 5 outlines the flow.Source: §4.3 · p. 17, Figure 5 · p. 18
4.4. Participants and Power Analysis
- The study recruited 455 participants, of whom 415 valid responses remained after 40 attention-check exclusions.The sample-size analysis estimated approximately 200 participants were needed across conditions.Source: §4.4 · p. 18
4.5. Measures
- Measures included EmCorp and GAToRS, with an additional explanation-satisfaction questionnaire for participants who saw explanations.The satisfaction scale adapted an eight-item, fully labelled five-point Likert questionnaire for robot explanations.Source: §4.5 · pp. 18–19, §4.5.1 · p. 19
4.5.1. Explanation Satisfaction Questionnaire
- Explanation satisfaction was assessed with an adapted eight-item scale rated on a five-point Likert scale.The scale was originally designed for explanations from software, algorithms, or tools.Source: §4.5.1 · p. 19
4.6. Study Procedure
- After baseline GAToRS and priming, participants viewed the failure video, completed EmCorp, and completed satisfaction measures if they received explanations.Post-GAToRS followed as the final step; both videos featured the same robot, and a semi-wizarded operation timed failures and explanations.Source: §4.6 · p. 19
5. Main Evaluation Results
In the main evaluation, explanations had a stronger positive impact for Furhat, especially under low expectations, while Pepper showed little change. By study end, general attitudes did not differ significantly across conditions.
5.1. Data Preparation
- The analysis excluded EmCorp’s mobility subscale; interpretation and expressiveness, GAToRS subscales, and explanation satisfaction had reported acceptable internal consistency.Cronbach’s alpha thresholds were at least 0.8 for interpretation and explanation satisfaction, at least 0.6 for expressiveness and societal-negative GAToRS, and at least 0.7 for the other GAToRS subscales.Source: §5.1 · p. 20
5.2. Baseline and Post-Study General Attitudes (GAToRS)
- Personal Negative baseline GAToRS scores differed across Furhat conditions (H(3) = 10.78, p = 0.012), but no GAToRS subscale differed significantly across conditions after the study.The post-study comparison found no significant condition differences on any subscale (p > 0.1).Source: §5.2 · p. 20
5.3. H1: Human Perception
- Figure 6 shows EmCorp ratings; Furhat expressiveness differed across conditions, and its interpretation scores were higher with explanations than without, while Pepper showed no significant modulation.For Furhat, expressiveness differed across conditions (H(3) = 15.73, p = 0.0012); explanation-versus-no-explanation interpretation differences were significant under both priming conditions.Source: Figure 6 · p. 21, §5.3 · pp. 21–22
5.4. H2: Explanation Satisfaction
- Figure 7 presents explanation-satisfaction results; Furhat participants were more satisfied after negative than positive priming, while Pepper showed no significant difference.The reported Furhat comparison was W = 994.5, p = 0.03; the corresponding medians were 3.56 ± 0.81 and 3.25 ± 0.60.Source: Figure 7 · p. 21, §5.4 · p. 22
- Explanations had a stronger positive impact when expectations were lower, particularly for Furhat, whereas Pepper showed no significant priming-related satisfaction difference.Furhat explanation satisfaction was higher after negative than positive priming (W = 994.5, p = 0.03; medians 3.56 ± 0.81 and 3.25 ± 0.60).Source: §5.4 · p. 22
6. Discussion
Explanations improved perceptions of Furhat, particularly among participants primed for low expectations, while Pepper showed no significant satisfaction improvement. The online, short-term studies and differing robot tasks constrain how broadly or durably these findings can be applied.
6. Discussion
- Positive and negative video priming induced high and low expectations, respectively, but priming effects could depend on the robot.The priming study showed strong expectation effects for both robots, while significant main-evaluation results appeared only for Furhat.Source: §6 · pp. 22–23
- Furhat explanations improved perceptions, with stronger positive effects when participants held lower expectations and explanations could shift blame from incompetence to technical constraints.Furhat participants primed negatively also rated its expressiveness significantly higher when it explained its mistake.Source: §6 · pp. 23–24
- Pepper explanations did not significantly improve satisfaction under either positive or negative priming.The authors suggest that Pepper’s movement and more disruptive physical errors may overshadow verbal explanations.Source: §6 · p. 24
6.1. Limitations
- The online video studies provide only a short-term snapshot, leaving real-world generalizability and the persistence of attitude changes unresolved.Attitudes were measured immediately after interactions; the authors call for more ecological or longitudinal studies.Source: §6 · pp. 22–23, §6.1 · p. 24
- The robots’ tasks and failures differed: Pepper’s physically disruptive errors contrasted with Furhat’s communication-based failures, potentially shaping responses to explanations.The online participant sample may also not represent society broadly in robot experience or interest.Source: §6.1 · p. 25
7. Future Work and Practical Takeaway
The authors recommend studying expectation management across contexts and aligning explanation strategies with robot embodiment and failure characteristics. Future work can test whether these findings persist in real-world interactions and under standardized tasks.
7. Future Work and Practical Takeaway
- Longitudinal or real-world studies could test whether short-term effects and explanation benefits persist with hands-on robot experience.The authors also propose comparing additional embodiments under standardized tasks.Source: §7 · p. 25
- Consistent failure scenarios and explanation strategies can help researchers examine context, while systematic comparisons of task types and failure severity may aid generalization.Pepper’s larger, more visible errors might overshadow explanations.Source: §7 · pp. 25–26
- Low-expectation users showed the largest gain from explanations, suggesting designers should attend to contexts where users may doubt robot competence from the outset.Healthcare is given as an example context.Source: §7 · p. 25
- Explanation modality should fit the robot’s communication style; Pepper’s movement and disruptive errors may call for integrated or multimodal explanations.The authors suggest combining speech with synchronized gestures or onscreen text, especially for mobile robots.Source: §7 · p. 25
8. Conclusion
The two-stage investigation validated video priming for high or low expectations across Pepper and Furhat, then found that explanations influenced perceptions especially when expectations were low. Different outcomes across the embodiments underscore the role of embodiment, task complexity, failure severity, and communication modality in explanation efficacy.
8. Conclusion
- Brief videos of capable versus failure-prone robot behavior reliably induced high or low expectations across Pepper and Furhat.The priming evaluation preceded the main evaluation.Source: §8 · p. 26
- Explanations significantly influenced user perception, especially when initial expectations were low.Source: §8 · p. 26
- Different outcomes for Pepper and Furhat highlight the interplay of embodiment, task complexity, failure severity, and communication modality in explanation efficacy.Source: §8 · p. 26
- The findings emphasize managing user expectations and tailoring explanation strategies to sustain trust and satisfaction when robots make mistakes.Source: §8 · p. 26