Paper deep dive
Back to the museum: Investigation of the acceptance of Android Andrea with and without emotion simulation in a museum
Marcel Heisler, Christian Becker-Asano
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/21/2026, 3:22:59 AM
Summary
This study investigates the acceptance of the android robot 'Andrea' in a German museum across three conditions: no emotion simulation, ChatGPT-based emotion simulation, and WASABI-based emotion simulation. Using a TAM2 questionnaire with 73 visitors, the results indicate that implementing emotion simulation (via ChatGPT or WASABI) did not yield positive effects on subjective evaluations compared to the non-emotional baseline. In fact, the non-emotional version was perceived as more useful and easier to use.
Entities (8)
Relation Signals (8)
Android Andrea ā deployedat ā Mercedes Benz Museum
confidence 95% Ā· A very human-like robot was installed in the Mercedes Benz Museum in Stuttgart, Germany
Android Andrea ā evaluatedby ā TAM2
confidence 90% Ā· An extended version of the Technology Acceptance Model Version 2 (TAM2) questionnaire was employed to let 73 visitors report on several factors of their opinion about the android robot Andrea
Android Andrea ā runson ā Nvidia Jetson Orin
confidence 90% Ā· The onboard software runs on a single Nvidia Jetson Orin
Android Andrea ā uses ā ChatGPT 4.1
confidence 90% Ā· the robots emotions were determined by ChatGPT 4.1
Android Andrea ā uses ā WASABI
confidence 90% Ā· the WASABI emotion simulation architecture simulated the robot's emotion dynamics
Android Andrea ā uses ā XTTSv2
confidence 85% Ā· Multilingual output is generated using XTTSv2 for speech synthesis.
Android Andrea ā uses ā FaceXHubert
confidence 85% Ā· To speed up this process, the ML model used previously is replaced by FaceXHubert
ChatGPT 4.1 ā comparedwith ā WASABI
confidence 80% Ā· Three experimental conditions were implemented, in which either (1) the robot simulated no emotions, (2) the robots emotions were determined by ChatGPT 4.1, or (3) the WASABI emotion simulation architecture simulated the robot's emotion dynamics.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:For a second time, the android robot Andrea was set up at a public museum in Germany for six consecutive days to have conversations with visitors, fully autonomously. Building on previously gathered qualitative results, the robot was now capable of engaging in multi-lingual conversation with the visitors about the museum context. The robot was prepared with context information about the museum in general and its surrounding exhibits this time. The robot featured a slightly artificial sounding voice that was previously evaluated as congruent with its gender-ambiguous but very humanlike design. Three experimental conditions were implemented, in which either (1) the robot simulated no emotions, (2) the robots emotions were determined by ChatGPT 4.1, or (3) the WASABI emotion simulation architecture simulated the robot's emotion dynamics. An extended version of the TAM2 questionnaire was employed to let 73 visitors report on several factors of their opinion about the android robot Andrea after having experienced it. In result, the statistical analysis suggests that these first two approaches to implementing emotions into the chat architecture of our android robot Andrea did not yield any positive effects on the subjective evaluations by the visitors and were not detectable on a conscious level.
Tags
Links
- Source: https://arxiv.org/abs/2607.16428v1
- Canonical: https://arxiv.org/abs/2607.16428v1
Trouble viewing inline? Open PDF directly ā
Full Text
35,597 characters extracted from source content.
Expand or collapse full text
Back to the museum: Investigation of the acceptance of Android Andrea with and without emotion simulation in a museum M. Heisler ā C. Becker-Asano ā heisler@hdm-stuttgart.de becker-asano@hdm-stuttgart.de Hochschule der Medien Stuttgart, Germany Figure 1: Android Andrea sitting on a bench in the museum and conversing with visitors Abstract For a second time, the android robot Andrea was set up at a pub- lic museum in Germany for six consecutive days to have conver- sations with visitors, fully autonomously. Building on previously gathered qualitative results, the robot was now capable of engag- ing in multi-lingual conversation with the visitors about the mu- seum context. The robot was prepared with context information about the museum in general and its surrounding exhibits this time. The robot featured a slightly artificial sounding voice that was previously evaluated as congruent with its gender-ambiguous but very humanlike design. Three experimental conditions were implemented, in which either (1) the robot simulated no emotions, (2) the robots emotions were determined by ChatGPT 4.1, or (3) the WASABI emotion simulation architecture simulated the robotās emo- tion dynamics. An extended version of the Technology Acceptance Model Version 2 (TAM2)questionnaire was employed to let 73 vis- itors report on several factors of their opinion about the android robot Andrea after having experienced it. In result, the statistical analysis suggests that these first two approaches to implementing emotions into the chat architecture of our android robot Andrea did not yield any positive effects on the subjective evaluations by the visitors and were not detectable on a conscious level. This work is licensed under a Creative Commons Attribution-NonCommercial- NoDerivatives 4.0 International License. IVA 2026, Puebla, Mexico Ā© 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2647-7/2026/09 https://doi.org/10.1145/3806774.3827982 CCS Concepts ā¢Human-centered computingāEmpirical studies in HCI; Empirical studies in interaction design. Keywords Human-Robot Interaction, empirical study, android robot, technol- ogy acceptance model ACM Reference Format: M. Heisler and C. Becker-Asano. 2026. Back to the museum: Investigation of the acceptance of Android Andrea with and without emotion simula- tion in a museum. InACM International Conference on Intelligent Virtual Agents (IVA 2026), September 07ā11, 2026, Puebla, Mexico.ACM, New York, NY, USA, 7pages.https://doi.org/10.1145/3806774.3827982 1 Introduction The intentional anthropomorphic design of android robots is meant to facilitate sophisticated multimodal communication. Consequently, they are presently being tested in scenarios where smooth, hu- manālike interaction is crucialāfor instance, to offer companion- ship and information to elderly users [ 5,25] or to serve in roles such as interviewers and receptionists [1,16]. In 2023, the android robot Andrea operated autonomously for six days at a German museum, engaging 44 visitors in unstructured conversations that later fed into structured interviews [ 11]. Visi- tors most frequently sought exhibit information, and interviewees highlighted faster response times and multilingual support as key improvements; genderāchanging cues had little effect on overall perception. arXiv:2607.16428v1 [cs.RO] 17 Jul 2026 IVA 2026, September 07ā11, 2026, Puebla, MexicoHeisler et al. Building upon these previous results, this paper presents a num- ber of technical improvements regarding the software architecture that drives the multimodal and multi-lingual behaviour of the an- droid robot. This includes the integration of two different approaches to simulate emotions that are expressed by the robot. In the aim to evaluate the impact of the emotion simulation on the visitorsā perception, the results derived from 73 questionnaires are being reported here. The remainder of this paper is structured as follows. In Sec- tion2work that is related is being discussed. Section3explains how the android robot was set up to answer visitor questions au- tonomously, before in Section4details of the questionnaire are pre- sented. The survey results are provided in Section5and discussed in Section6. 2 Related work Before the availability of large language models such as ChatGPT, natural language processing for interactive, autonomous, social robots has been a major challenge [ 13]. Most early social robots did not feature a very human-like design, but a mobile platform enabling them to guide visitors through the museum. Android robots are a specialized subset of the broader category of social robots, many of which have been deployed in museum set- tings. A review of social robots in museums [8] identified that the predominant role of these machines is to act as guides; they are less frequently used for entertainment or educational purposes. The study also found that the majority of such robots are deemed ac- ceptable for museum environments and distilled three key themes from the literature that underpin successful humanārobot interac- tion in these contexts, namely appropriate facial expressions, move- ment, and communication abilities. In Linz, Austria, the android Geminoid HIā1 was positioned be- hind a table in the public cafĆ© atop the Ars Electronica museum, accompanied by informational panels about Japan [ 24]. Three op- erating modes were tested: (1) passively monitoring a laptop placed in front of it, (2) autonomously making eye contact with passersby, and (3) being teleāoperated. Structured interviews and video analy- ses revealed that increased eye contact made the robot more readily identifiable as a machine, while participants tended to find it more interesting than unsettling. Later, Geminoid HIā1 was teleāoperated during the Ars Electron- ica festival to engage with visitors [ 3]. Qualitative interviews indi- cated a predominance of positive over negative remarks about the robot; 37.5% of respondents reported feeling uncanny, whereas 29% enjoyed conversing with it. To probe the Uncanny Valley phenomenon, museum patrons interacted with a teleāoperated Telenoid robot that had been in- stalled in the Ars Electronica Centerās public robot laboratory in 2015 [ 21]. Unlike most androids, this robotās look is far less hu- manālike. The study found that when visitors were first given a sci- enceāfiction narrative framing the encounter, the robotās eeriness ratings dropped noticeably. Nadine the Social Robotwas housed at Singaporeās ArtScience Museum in 2017 for a fiveāmonth period, but no formal user study could be conducted because of a confidentiality clause [ 25]. The same android robot was recently tested in a semi-controlled setup Figure2:Theandroidrobotplacedonabenchinthemuseum with a microphone for verbal interaction inside the lobby of the University of Geneva [14]. The question- naire results suggest that an LLMāpowered hyperārealistic robot enhances usersā perceptions of pleasantness and approachability while slightly reducing creepiness, with conversational naturalness and interestākeeping being the primary drivers of engagement. It recommends prioritizing fluid, engaging dialogue over mere hu- manālike mimicry and calls for more diverse, objective, and longāterm studies to confirm these findings. Research on social humanoid robots such as Ameca [ 4] demon- strates that the uncanny valley effect depends on a delicate inter- play between appearance, emotional expression, and interaction skills. These findings suggest that balancing humanālike features with functional behavior and adaptive emotion can improve usabil- ity and mitigate uncanny feelings. Ameca robots are also exhibited in various museums, like Heinz Nixdorf MuseumsForum 1 , Zukun- ftsmuseum NĆ¼rnberg 2 or Museum of the Future, Dubai 3 , but no formal user study evaluating the presence of the robot was con- ducted. In 2023, an autonomous run of six days saw the android Andrea deployed in a German museum, where it conversed freely with 44 visitors [11]. The dialogue data revealed that most guests used Andrea primarily to inquire about exhibit details, and the analysis results of structured interviews suggest that visitors find quicker replies and multilingual capability to be necessary upgrades. No- tably, variations in genderāchanging cues did not markedly influ- ence how the robot was perceived. This previous work shows the need to research the effectiveness and acceptance of more emotionally interactive, android robots employed in public spaces such as a museum. 3 Technical setup A very human-like robot was installed in the Mercedes Benz Mu- seum in Stuttgart, Germany, for six consecutive days freely acces- sible for the visitors. The android robot was seated on a bench in 1 https://w.hnf.de/dauerausstellung/ausstellungsbereiche/global-digital/mensch- roboter-leben-mit-kuenstlicher-intelligenz-und-robotik/ameca.html 2 https://w.deutsches-museum.de/nuernberg/aktuell/roboter-ameca-im- zukunftsmuseum 3 https://gulfnews.com/uae/science/dubai-museum-of-the-futures-upgraded-ai- powered-humanoid-robot-speaks-hindi-too-1.500104771 Back to the museumIVA 2026, September 07ā11, 2026, Puebla, Mexico Figure 3: Overview of the software architecture of the conversational functions of the android Andrea. one of the main exhibition halls of the museum, called Mythos 6, with experimental cars of the museum ownerās brand surrounding it. The android robot was assembled by the Japanese firm AāLab. Its sitting posture is comparable to the companyās Premium Model, but it features additional actuators for independent control of each eyelid and finger. Altogether, 52 pneumatic actuators manage the upper body and facial movements. The robotās exterior is deliberately androgynous, and a similarly androgynous voice was designed for it in previous research [ 20]. It wears casual blue jeans and a gray zipāup hoodie emblazoned with a small university logo (see Fig.1and2). Cameras are mounted in both of the robotās eyes to capture vi- sual input. Meanwhile, conversational interaction is facilitated by a hidden speaker concealed inside the hoodie and a microphone placed on a small table directly in front of the android. The on- board software runs on a single Nvidia Jetson Orin housed in the bench under the robot. Power is supplied by electricity and com- pressed air (7 bar), while data communication with the Jetson is established via a USB connection. The compressor is placed inside a custom-build, black counter desk that is placed approximately five meters behind the back of the robot and is covered with con- voluted acoustic foam inside to dampen the noise. The software components are an extended version of [ 10,11], which combines four separate, open-sourceMachine Learning (ML) models and OpenAIās ChatGPT assistant function, see also Fig.3: (A)ForSpeech-to-Text (STT)whisper-large-v3-turbois used supporting multiple languages as input. The input text is then transmitted to an OpenAI assistant. (B)The user request is taken as input for a ChatGPT4.1-based OpenAI-assistant that is equipped with a special system prompt and uses Retrieval Augmented Generation (RAG) as described below. Also, the three different experimental conditions are realized in this step. (C)Multilingual output is generated using XTTSv2 [6] for speech synthesis. This ML model supports 17 languages. A specific design-congruent voice [19] was cloned and used consis- tently in all languages. (D)The generation of lip-sync facial animation is described in [12]. To speed up this process, the ML model used previously is replaced by FaceXHubert [9]. (E)The people tracking is based on a realtime analysis of the video stream provided by the camera in the robotās left eye usingposenet[18] (not shown in Fig.3). For initiating a dialogue the user must manually enable the mi- crophone, which triggers a listening animation. We deliberately avoided passive listening to reduce the likelihood of spurious ac- tivations in noisy public spaces. Once the microphone is disabled, the robot switches to a āthinkingā animation, signaling that it is processing the userās input. As shown in [22], such thinking faces improve the naturalness of humanārobot interaction. To further humanize its behavior, the robot blinks at random intervals and yawns when no interaction has occurred for a while. To generate context-sensitive textual responses, an OpenAI as- sistant was used withgpt-4.1as the basis. The assistantās prompt contains information about the robot itself and about its surround- ing, e.g., which cars are located where. In addition, RAG is used to provide additional information, such as the transcripts of all audio guide data in English. When the robotās emotional states are simulated using WASABI [ 2], first, the OpenAI assistant is prompted to calculate the valence of the last input sentence spoken by the user. The resulting value ranges from -100 to 100 and is sent to WASABI as a valenced im- pulse. In effect, WASABI, which is running as a concurrent process IVA 2026, September 07ā11, 2026, Puebla, MexicoHeisler et al. in the local computer, returns the emotion likelihood of the emo- tions āhappyā, āsadā, āangryā, āfearfulā, āsurprisedā or āneutralā, of which the one with the highest value is used to animate the robotās face asynchronously. The internal emotion dynamic of WASABI lets the robotās emotional state automatically return to āneutralā after some time without any inputs. Alternatively, ChatGPT is in- structed to generate one of these emotion labels directly. The inten- sity dynamics of these emotions is then taken care of in the main process of the software architecture by employing a linear decay function. All emotions are expressed using validated static facial expres- sions [15]. The animations for thinking and speech have a higher priority than the emotional expressions for the relevant actuators. Thus, while the robot is speaking, its mouth movements synchro- nize with the speech signal; before or after speaking, however, the facial actuators reflect the simulated emotion (e.g., smiling). 4 Method Due to the circumstances inside a public museum, the study had to follow a between-groups design. On the first and fourth day the ro- bot was in the āno emotionā condition, on the second and fifth day it was in the āChatGPTemotionā emotional condition, and on the third and sixth day it was in the āWASABIā emotional condition. However, due to hardware issues the robotās facial expressions had degraded so much on the sixth day that the questionnaire data of that day had to be excluded from the analysis. 4.1 Study procedure On each day, a team of three experimenters was present on-site in alternating shifts to supervise the study setup and ensure a safe environment. Depending on the situation and the experimenter on duty, interested visitors were at times actively approached, while at other times they initiated the interaction themselves. If necessary, the experimenters provided a brief introduction on how to interact with the android robot and explained the use of the microphone. After the interactions, users were invited to complete a question- naire that contained an informed consent and items in both English and German. Additionally, participants were informed that they could cancel the survey at any time. In total,ķ = 73visitors vol- untarily provided questionnaire data, withķ ķ¶āķķ”ķŗķķķķķķ”ķķķ = 35, ķ ķ¶āķķ”ķŗķķķķ¢ķķ = 24andķ ķ ķ“ķķ“ķµķ¼ = 14, cf. Table 1. The TAM2 questionnaire from [23] was used as a basis, as it provided the exact German translations of the original items [26]. Our questionnaire can be divided into the following categories: general information such as date, time, and age; questions eval- uating participantsā prior knowledge of the robotās availability at this place and whether it influenced their decision to visit the mu- seum; TAM2 items measuring intention to use (ITU),perceived use- fulness (PU), andperceived ease of use (PEOU); questions regarding the emotions; as well as additional questions, regarding for how long participants had interacted with the robot, which languages they used, and any suggestions they had. Finally, participants were asked how useful they considered the robot in the museum con- text. Table 1: Descriptives of the questionnaire results Shapiro-Wilk Conditionķķķķ ķķ· ķķ ITUChatGPTemotion4.99 1.67 0.88.001 ITUChatGPTpure5.69 1.51 0.82 < .001 ITUWASABI4.93 1.73 0.91.172 PUChatGPTemotion4.19 1.57 0.95.115 PUChatGPTpure5.21 1.74 0.87.005 PUWASABI4.18 1.90 0.92.230 PEOU ChatGPTemotion5.08 1.18 0.91.010 PEOU ChatGPTpure5.89 0.83 0.89.013 PEOU WASABI5.32 1.69 0.83.011 UTChatGPTemotion7.53 2.72 0.84 < .001 UTChatGPTpure7.88 2.11 0.87.005 UTWASABI7.71 2.81 0.79.004 4.2 Measurement Whether participants knew about the robotās availability prior to their visit, and whether they had specifically come to see the robot was evaluated with yes/no questions. The TAM2 items and emotion-related questions were evaluated using a 7-point Likert scale, where 1 indicated āstrongly disagreeā and 7 indicated āstrongly agreeā. āIntention to Useā was measured with two items: (1)Assuming I have access to the robot, I intend to use it. (2)Given that I have access to the robot, I predict that I would use it. āPerceived Usefulnessā was measured with four items: (1)Using the robot improves my performance. (2)Using the robot increases my productivity. (3)Using the robot enhances my effectiveness. (4)I find the robot to be useful. āPerceived ease of Useā was measured with four items: (1)My interaction with the robot is clear and understandable. (2)My interaction with the robot does not require a lot of my mental effort. (3)I find the robot to be easy to use. (4)I find it easy to get the robot to do what I want it to do. The Shapiro-Wilk test was used to determine theķstatistic andķ- value, in order to assess whether the data is normally distributed, see Table1. The perceived emotionality of the robot was measured with three items: (1)I have the impression that the robot displays emotional re- actions (2)I find the robotās emotional responses appropriate to the sit- uation. (3)The robot appears moody or shows emotional mood swings. The final question āHow useful do you think would it be to use this robot here today?ā (reported here as UsefulnessToday (UT)) was measured by a distinct Likert scale ranging from 0 (ānot at allā) Back to the museumIVA 2026, September 07ā11, 2026, Puebla, Mexico to 10 (āvery muchā) to compare our setup with a previous result of a study with the robot in the same public museum [11]. 5 Results Only five of all 73 interviewees knew about the presence of the robot beforehand and only one of them came explicitly to interact with it on that day. Nineteen interviewees were in the youngest age range between 10 and 20 years old (ChatGPTpurefour,ChatGPTe- motion13, andWASABItwo), 22 were between 21 and 30 years of age (ChatGPTpurefour,ChatGPTemotion12, andWASABIsix), and the remaining 25 were older than 30 years. The robot was spoken to in German 788 times and in English 659 times, followed by Turk- ish (102), Spanish (82), and Russian (61) times. It responeded most often in German by producing a total of 2023 German sentences, followed by English (1212), Turkish (324), Spanish (194), and Rus- sian (162). When the input sentence was very short, the language detection failed sometimes and the robot answered in a different language, though. An overview of the remaining questionnaire results is presented in Figures 4-7. These box plots show that those visitors, who had interacted with the non-emotional version of the android robot (ChatGPTpure), found it most useful (Figure5), easiest to use (Fig- ure6) and had the highest intention to use it (Figure4). They also did not find the robot less useful today as compared to the groups of visitors who had interacted with one of the emotional variants (Figure7). To test these results for significance 4 , normality of the data was evaluated using the ShapiroāWilk test for each variable and exper- imental condition, cf. Table1. For several groups the test indicated significant deviations from a normal distribution (ķ < 0.05, for ITU, PU and UT). As the assumption of normality was not fully met in all dimensions, a non-parametric KruskalāWallis test was applied to compare the experimental conditions. Table 2: Kruskal-Wallis test results ķ 2 ķķķ ķ 2 IntentionToUse3.552 .169 0.05 PerceivedUsefulness5.762 .056 0.08 PerceivedEaseOfUse6.832 .033 0.09 How useful do you think ...? 0.132 .939 0.00 The obtainedķ-values presented in Table2for PU and PEOU remain close to or even below the desired level of significance (ķ¼ = 0.05) with a moderate effect size (ķ > 0.6). Accordingly, pair- wise comparisons were applied using the Dwass-Steel-Critchlow- Flinger test separately for these two dimensions. The results reveal that in both cases only the difference betweenChatGPTemotion andChatGPTpureis significant, cf. Table 3and Table4. The manipulation check whether the different visitors perceived the emotional expressions in the two conditionsChatGPTemotion andWASABIis presented in Table 5. It can be seen that all averages lie close to four, which is the center point on the seven point likert 4 All tests were performed using theJAMOVIopen source software (version 2.7.24.0). Table 3: Pairwise comparisons - PerceivedUsefulness ķķ ChatGPTemotion ChatGPTpure 3.39 .044 ChatGPTemotion WASABI-0.14 .995 ChatGPTpureWASABI-2.11 .296 Table 4: Pairwise comparisons - PerceivedEaseOfUse ķķ ChatGPTemotion ChatGPTpure 3.63 .028 ChatGPTemotion WASABI2.01 .329 ChatGPTpureWASABI-0.77 .848 Table 5: Descriptives of the three questions regarding the emotionality of the robot: (1) I have the impression that the robotdisplaysemotionalreactions,(2)Ifindtherobotāsemo- tional responses appropriate to the situation, (3) The robot appears moody or shows emotional mood swings. Q. Conditionķ ķķķķ ķķķķķķ ķķ· (1) ChatGPTemotion 354.404.00 1.83 (1) ChatGPTpure244.044.00 1.99 (1) WASABI143.503.00 1.87 (2) ChatGPTemotion 354.314.00 1.75 (2) ChatGPTpure244.544.50 2.11 (2) WASABI144.434.00 1.99 (3) ChatGPTemotion 353.313.00 2.01 (3) ChatGPTpure242.922.50 1.74 (3) WASABI143.003.00 1.92 scale. Also, there is reason to assume that the values are not dis- tributed normally according to lowķ-values of the Shapiro-Wilk tests. The results of the Kruskal-Wallis test (cf. Table6) show no significant differences between conditions. Table 6: Kruskal-Wallis test results for perceived emotional- ity / mood swings: (1) I have the impression that it displays emotionalreactions,(2)Ifinditsemotionalresponsesappro- priatetothesituation,(3)Therobotappearsmoodyorshows emotional mood swings. Q.ķ 2 ķķķ (1) 2.592 .274 (2) 0.292 .866 (3) 0.512 .773 6 Discussion and conclusion To the best of our knowledge, this work is the first to investigate the impact of emotional reactions of an interactive, autonomous, android robot in a public space. Unfortunately, the results seem IVA 2026, September 07ā11, 2026, Puebla, MexicoHeisler et al. 2 4 6 ChatGPTemotionChatGPTpureWASABI Experimental Condition IntentionToUse Figure 4: Box plot showing the distribution ofintention to use (ITU)by experimental condition with indicated median values. 2 4 6 ChatGPTemotionChatGPTpureWASABI Experimental Condition PerceivedUsefulness Figure 5: Box plot showing the distribution ofperceived use- fulness (PU)by experimental condition with indicated me- dian values. 2 4 6 ChatGPTemotionChatGPTpureWASABI Experimental Condition PerceivedEaseOfUse Figure 6: Box plot showing the distribution ofperceived ease of use (PEOU)by experimental condition with indicated me- dian values. 0.0 2.5 5.0 7.5 10.0 ChatGPTemotionChatGPTpureWASABI Experimental Condition How useful do you think ...? Figure 7: Box plot showing the distribution of theUseful- nessToday (UT)by experimental condition with indicated median values on a scale from zero to ten. to indicate that both of our ways to implement an emotion pro- cess that derives emotions from user input and lets the robot ex- press them via its face seem not to enhance its overall acceptance, but rather diminish it, although only to a very small extend. The negative result of the manipulation check further devalues our ap- proach. One possible reason for this may be the very public setup, which can keep people from engaging in deeply emotional conver- sations. In fact, inspecting the logs retrospectively revealed that emotions other than āhappyā have hardly been displayed at all. While the quantitative results of the survey remain somewhat inconclusive, qualitative observations made by the experimenters during the study provide valuable context regarding human-robot interaction in the wild. Notably, the robotās highly human-like ap- pearance effectively camouflaged it within the museum environ- ment. Many visitors initially mistook Andrea for a regular museum patron. The realization that it was an android often occurred only upon hearing it speak or after consulting the experimenters. Furthermore, as noted in the study procedure, the initial thresh- old for engaging with the robot proved to be relatively high. The frequent need to actively encourage patrons and explain the micro- phoneās usage highlights that spontaneous interaction with such systems is not yet fully intuitive. Once approached, visitors of- ten required guidance on suitable conversation topics. Interactions typically commenced with cautious exploration, where users would ātest the watersā by asking museum-related questions. However, as users became more accustomed to the system, the dialogue fre- quently expanded into an array of broader subjects, ranging from sports and the robotās technical capabilities to philosophical in- quiries about the meaning of life. Language expectations also played a significant role in the visi- torsā initial hesitation, as many assumed the android was restricted to German or English. When the experimenters assured them that Andrea could comprehend and reply in their native languages, vis- itors exhibited visible satisfaction and a greater willingness to con- verse. This observation highlights the impact that accessible, multi- lingual capabilities have on fostering inclusive and engaging inter- actions in public spaces. However, our study suffers from several limitations. With only ķ = 14questionnaire results the WASABI condition is likely un- derpowered and small effects remain undetectable. As explained, technical problems on the last day did not allow us to collect more data in this condition as we had planned initially. The robotās ges- ture behavior was rather limited as well. Apart from rising its left arm to its ear, making a fist with its left hand while looking down āin thoughtsā, and moving the arm back down again when start- ing to speak, Andrea did not use any iconic, metaphoric or even manipulative gestures. Furthermore, the robot displays emotions only unimodally through facial expressions, while humans rely on additional modalities, like body posture and voice cues, to decode them [ 17]. Back to the museumIVA 2026, September 07ā11, 2026, Puebla, Mexico In our current work, first, we aim to improve the visibility of the emotional states by modulating the robotās speech according to the activated emotion at runtime dynamically. This can be achieved by deriving embedding vectors that represent emotions in the speaker embedding space of aText-to-Speech system as, for example, pro- posed by theEmoKnob[7] framework. Secondly, a personalized autobiographical memory has recently been implemented in the aim to let robot Andrea remember and reuse person-specific, con- versational knowledge in its dialogue in long-term, repeated inter- actions. Extending the gesture behavior by integrating automated function calling in the LLM-pipeline is also an open task for future engineering work. Acknowledgments We would like to warmly thank the Mercedes-Benz Museum staff for their continuous support during the setup and the entire six- day exhibition, the students Leon Kiefer, Steve Aschenbrenner, and Dilara Celepci-Uludag for their help, as well as all museum visitors who kindly participated in our study and provided their feedback. References [1]Evangelia Baka, Nidhi Mishra, Emmanouil Sylligardos, and Nadia Magnenat- Thalmann. 2022. Social Robots and Digital Humans as Job Interviewers: A Study of Human Reactions Towards a More Naturalistic Interaction. InHuman- Computer Interaction. Technological Innovation, Masaaki Kurosu (Ed.). Springer International Publishing, Cham, 455ā474. doi:10.1007/978-3-031-05409-9_34 [2]C. Becker-Asano. 2014. WASABI for affect simulation in human-computer in- teraction. InProc. on Emotion Representations and Modelling for HCI Systems. Springer, Sydney, Australia. [3]Christian Becker-Asano, Kohei Ogawa, Shuichi Nishio, and Hiroshi Ishiguro. 2010. Exploring the uncanny valley with Geminoid HI-1 in a real-world ap- plication. InProceedings of IADIS International conference interfaces and human computer interaction. 121ā128. [4]Karsten Berns and Ashita Ashok. 2024. āYou Scare Meā: The Effects of Humanoid Robot Appearance, Emotion, and Interaction Skills on Uncanny Valley Phenom- enon.Actuators13, 10 (Oct. 2024), 419. doi:10.3390/act13100419 [5]Felix Carros, Berenike BĆ¼rvenich, Ryan Browne, Yoshio Matsumoto, Gabriele Trovato, Mehrbod Manavi, Keiko Homma, Toshimi Ogawa, Rainer Wieching, and Volker Wulf. 2022. Not that Uncanny After All? An Ethnographic Study on Android Robots Perception of Older Adults in Germany and Japan. InSocial Robotics. Vol. 13818. Springer Nature Switzerland, Cham, 574ā586. doi:10.1007/ 978-3-031-24670-8_51Series Title: Lecture Notes in Computer Science. [6]Edresson Casanova, Kelly Davis, Eren Gƶlge, Gƶrkem Gƶknar, Iulian Gulea, Lo- gan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, and Ju- lian Weber. 2024. XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model. InInterspeech 2024. 4978ā4982. doi:10.21437/Interspeech.2024-2016 [7]Haozhe Chen, Run Chen, and Julia Hirschberg. 2024. EmoKnob: Enhance Voice Cloning with Fine-Grained Emotion Control. arXiv: 2410.00316[cs.CL]https: //arxiv.org/abs/2410.00316 [8]Norina Gasteiger, Mehdi Hellou, and Ho Seok Ahn. 2021. Deploying social robots in museum settings: A quasi-systematic review exploring purpose and ac- ceptability.International Journal of Advanced Robotic Systems18, 6 (Nov. 2021), 17298814211066740. doi:10.1177/17298814211066740Publisher: SAGE Publica- tions. [9]Kazi Injamamul Haque and Zerrin Yumak. 2023. FaceXHuBERT: Text- less Speech-driven E(X)pressive 3D Facial Animation Synthesis Using Self- Supervised Speech Representation Learning. InProceedings of the 25th Interna- tional Conference on Multimodal Interaction (ICMI ā23). Association for Comput- ing Machinery, New York, NY, USA, 282ā291. doi:10.1145/3577190.3614157 [10]Marcel Heisler and Christian Becker-Asano. 2023. An Android Robot Head as Embodied Conversational Agent. InISR Europe 2023; 56th International Sympo- sium on Robotics. 93ā99.https://ieeexplore.ieee.org/document/10363058 [11]Marcel Heisler and Christian Becker-Asano. 2025. Conversations with An- drea: Visitorsā Opinions on Android Robots in a Museum. In2025 34th IEEE International Conference on Robot and Human Interactive Communication (RO- MAN). IEEE, Eindhoven, Netherlands, 112ā119. doi:10.1109/RO-MAN63969. 2025.11217762 [12]Marcel Heisler, Stefan Kopp, and Christian Becker-Asano. 2023. Making an Android Robot Head Talk. In2023 32nd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). 1837ā1842.doi:10.1109/RO- MAN57019.2023.10309532ISSN: 1944-9437. [13]Mehdi Hellou, JongYoon Lim, Norina Gasteiger, Minsu Jang, and Ho Seok Ahn. 2022. Technical Methods for Social Robots in Museum Settings: An Overview of the Literature.Int J of Soc Robotics14, 8 (Oct. 2022), 1767ā1786.doi:10.1007/ s12369-022-00904-y [14]Hangyeol Kang, Thiago Freitas, Maher Ben Moussa, and Nadia Magnenat Thal- mann. 2026. Affective and Conversational Predictors of Re-Engagement in HumanāRobot Interactions: A Student-Centered Study with a Humanoid Social Robot.Int J of Soc Robotics18, 3 (March 2026), 41.doi:10.1007/s12369-026-01377- z [15]Amelie Kassner and Christian Becker-Asano. 2023. Comparing an android head with its digital twin regarding the dynamic expression of emotions. In2023 11th International Conference on Affective Computing and Intelligent Interaction Work- shops and Demos (ACIIW). [16]Tatsuya Kawahara, Koji Inoue, and Divesh Lala. 2021. Intelligent Conversational Android ERICA Applied to Attentive Listening and Job Interview.http://arxiv. org/abs/2105.00403 [17]Dacher Keltner and Daniel T. Cordaro. 2017. Understanding Multimodal Emo- tional Expressions: Recent Advances in Basic Emotion Theory. InThe Science of Facial Expression, James A. Russell and Jose Miguel Fernandez Dols (Eds.). Ox- ford University Press.doi:10.1093/acprof:oso/9780190613501.003.0004 [18]Alex Kendall, Matthew Grimes, and Roberto Cipolla. 2015. PoseNet: A Convo- lutional Network for Real-Time 6-DOF Camera Relocalization. In2015 IEEE Intl. Conf. on Computer Vision (ICCV). 2938ā2946. doi:10.1109/ICCV.2015.336ISSN: 2380-7504. [19]Johanna Magdalena Kuch, Marcel Heisler, Stina Klein, Silvan Mertes, Lennart Eing, Elisabeth AndrĆ©, and Christian Becker-Asano. 2025. Your Robot, My Voice: Enhancing Android Robot Likability through Personalization by Cloning the Userās Voice. In2025 34th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). IEEE, Eindhoven, Netherlands, 192ā198. doi:10.1109/RO-MAN63969.2025.11217611 [20]Johanna Magdalena Kuch, Jauwairia Nasir, Silvan Mertes, Ruben Schlagowski, Christian Becker-Asano, and Elisabeth AndrĆ©. 2024. Evaluating Gender Ambigu- ity, Novelty and Anthropomorphism in Humming and Talking Voices for Robots. In2024 33rd IEEE International Conference on Robot and Human Interactive Com- munication (ROMAN). 2219ā2225. doi:10.1109/RO-MAN60168.2024.10731423 ISSN: 1944-9437. [21]Martina Mara and Markus Appel. 2015. Science fiction reduces the eeriness of android robots: A field experiment.Computers in Human Behavior48 (July 2015), 156ā162.doi:10.1016/j.chb.2015.01.007 [22]Shushi Namba, Wataru Sato, Saori Namba, Alexander Diel, Carlos Ishi, and Takashi Minato. 2024. How an Android Expresses āNow Loading...ā: Examin- ing the Properties of Thinking Faces.Int J of Soc Robotics16, 8 (Aug. 2024), 1861ā1877. doi:10.1007/s12369-024-01163-9 [23]Thomas Olbrecht. 2010.Akzeptanz von E-Learning. Eine Auseinandersetzung mit dem Technologieakzeptanzmodell zur Analyse individueller und sozialer Ein- flussfaktoren.Univ., Jena. 215 S. pages. http://w.db-thueringen.de/servlets/ DerivateServlet/Derivate-21996/Olbrecht/Dissertation.pdf [24]Astrid M Rosenthal-von der PĆ¼tten, Nicole C KrƤmer, Christian Becker-Asano, Kohei Ogawa, Shuichi Nishio, and Hiroshi Ishiguro. 2014. The uncanny in the wild. Analysis of unscripted humanāandroid interaction in the field.Interna- tional Journal of Social Robotics6 (2014), 67ā83. [25]Nadia Magnenat Thalmann, Nidhi Mishra, and Gauri Tulsulkar. 2021. Nadine the Social Robot: Three Case Studies in Everyday Life. InSocial Robotics. Springer International Publishing, Cham, 107ā116.doi:10.1007/978-3-030-90525-5_10 [26]Viswanath Venkatesh and Fred D. Davis. 2000. A Theoretical Extension of the Technology Acceptance Model: Four Longitudinal Field Studies.Management Science46, 2 (2000), 186ā204.