Source-linked AI summary
From Small Talk to Rapport: Exploring Robot Self-Disclosure in Collaborative Tasks
Kaitlynn Taylor Pineda, Anvii Mishra, Brian Chien, Angela Guo, Toluwani Williams, Ziang Xiao, Chien-Ming Huang
TL;DR
Designing effective small talk for human–robot collaboration remains unclear, especially for non-anthropomorphic robots. This study compared low- and high-disclosure strategies in an LLM-driven manipulator with 50 participants and found that low disclosure elicited more participant disclosure and stronger teaming and coordination.
Problem
It remains unclear which content and style of robot-initiated small talk best support human–robot team dynamics.
Method
An LLM-driven non-anthropomorphic manipulator used low- or high-disclosure small talk during a cooperative task with 50 participants.
Results
Participants disclosed more and reported stronger teaming and coordination with the low-disclosure robot, while prior robot experience amplified lower receptivity to high disclosure.
Takeaways & Limitations
Robot self-disclosure should be calibrated to context and user expectations because greater disclosure may hinder reciprocity and human–robot teamwork.
Takeaways & Limitations
The simulated quality-control task, largely university-based sample, fixed disclosure details, and single non-humanoid robot limit generalizability.
Abstract
from arXiv · showhide
People naturally chat while collaborating and share personal information (i.e., self-disclose) to build rapport and maintain social connections. As robots are increasingly developed to work with people, the effective use of these social behaviors to enhance engagement and support teamwork becomes ever more important. While prior work has shown that robot-initiated small talk can benefit human-robot collaboration, less is known about how best to design such small talk. In this work, we explore how self-disclosure may be designed to support small talk within a human-robot team---especially when the robot is an industrial manipulator that lacks anthropomorphic cues and performs physical work. We first developed an LLM-driven manipulator capable of partaking in small talk, adopting either a low-disclosure or high-disclosure strategy. We then conducted a user study (N = 50) to investigate how self-disclosure in small talk influences human-robot dynamics. Unexpectedly, participants disclosed more in the low-disclosure condition and reported stronger teaming and coordination than those in the high-disclosure condition. This effect was more pronounced among users with prior experience teaming with robots. These results suggest that increasing robot self-disclosure does not necessarily foster rapport, social connection, or reciprocal disclosure; other factors, such as prior HRI experience, should be considered.
I. INTRODUCTION … B. Self-Disclosure in Dialogue Systems
The paper examines how self-disclosure-oriented small talk affects collaboration with non-anthropomorphic manipulators, addressing uncertainty about the content and style of robot small talk. It builds on evidence that small talk and disclosure can support engagement, rapport, trust, and reciprocal sharing, while testing whether these effects generalize to industrial robots.
- I. INTRODUCTION: Non-anthropomorphic robots are widely deployed in manufacturing and logistics to perform physically demanding or precision tasks alongside people.Their common use makes social communication relevant to practical human–robot collaboration.
- I. INTRODUCTION: Robot-initiated small talk can improve people’s engagement and rapport with non-anthropomorphic robots compared with no small talk.However, the most effective small-talk content and style for team dynamics remain unclear.
- I. INTRODUCTION: Self-disclosure is the voluntary sharing of personal thoughts, feelings, or information and supports rapport and trust,.Prior HRI studies also report that disclosure can elicit reciprocal disclosure and strengthen perceived social bonds,.
- B. Self-Disclosure in Dialogue Systems: It remains unknown whether disclosure effects observed with humanoid or social robots generalize to non-anthropomorphic manipulators.Although small talk benefits collaboration with such manipulators, how people respond to and perceive disclosure-oriented small talk remains unresolved.
- I. INTRODUCTION: The study evaluates an autonomous LLM-driven manipulator with low- versus high-disclosure small-talk strategies in a co-working scenario involving 50 participants.This design examines partners’ willingness to self-disclose and their perceptions of the robot during collaboration.
- A. Self-Disclosure Through Small Talk: Social penetration theory proposes that self-disclosure gradually increases as relationships develop, building closeness and trust.Small talk offers a low-stakes setting for incremental and reciprocal sharing that can deepen from superficial information over time,.
- B. Self-Disclosure in Dialogue Systems: Dialogue-system studies find that emotional disclosures can encourage reciprocal sharing, increase enjoyment, and improve perceptions of agents,,.Participants have also shared highly intimate or negative disclosures more readily with robots than humans, possibly because robots seem non-judgmental.
III. ROBOT SYSTEM
The robot system combined speech recognition, a GPT-4o conversational pipeline, and a 7-DOF Franka Emika Panda manipulator. Conversation was generated separately from pre-programmed physical trajectories for sorting PVC pipes, with helper LLM agents supporting the interaction workflow.
- System Architecture: The system integrated ASR, GPT-4o conversation generation, and a 7-DOF Franka Emika Panda manipulator.The conversational pipeline handled spoken interaction, while physical behavior was executed separately through pre-programmed motion trajectories.
- Physical Task: During each sub-task cycle, the robot picked up, inspected, and sorted PVC pipes into yellow, green, or faulty locations before returning home.
- Conversational Workflow: The conversational workflow used multiple helper LLM agents to listen, decide whether to respond, and refine generated replies.Figure 2 illustrates the workflow and provides a detailed example for a high-disclosure participant.
A. Speech Listener: Automatic Speech Recognition (ASR) … A. Study Task
The system combines chunked Google Cloud ASR with an LLM conversation workflow that manages turn-taking, response generation, questioning, and condition-specific disclosure. The study uses a loosely coupled factory task in which humans and robots divide quality-control labor while the robot initiates chit-chat.
- A. Speech Listener: Automatic Speech Recognition (ASR): The ASR stream transcribes speech with Google Cloud Speech-to-Text in 2.6-second chunks at 16 kHz, using silence to mark response completion.The ASR pauses while the LLM processes input and the robot speaks; this setup is similar to prior work.
- B. LLM-based Conversation Generation: Transcribed speech enters a conversational workflow of multiple LLM-powered components, with complete prompts provided in Appendix E3.The workflow is illustrated in Fig. 2.
- B. LLM-based Conversation Generation: A Filter Agent classifies user speech as Wait, Respond, or End so the workflow advances only when input is ready.Wait indicates that a response is still in progress, Respond passes completed input downstream, and End indicates that no reply is needed.
- B. LLM-based Conversation Generation: A response-generation component tailors candidate utterances to user input, configuring an extroverted tone, concise length, and alignment with participant affect.The candidate response serves as a base utterance for downstream refinement.
- B. LLM-based Conversation Generation: A Sentence Rephraser Agent limits excessive robot questioning by rewriting candidate responses into non-questions when questions exceed 50%.This component is intended to promote balanced conversational flow by monitoring robot question frequency.
- B. LLM-based Conversation Generation: In the high-disclosure condition, a Disclosure Rewriter Agent enables explicit robot self-disclosure.The supplied passage describes this condition-specific rewriting component but does not provide further implementation details.
- A. Study Task: The cooperative task assigned the robot to sort PVC pipes and the participant to inspect and sort PVC fittings toward a shared quality-control goal.The task was deliberately loosely coupled rather than synchronous joint manipulation, while the robot initiated chit-chat during sorting.
B. Study Conditions … V. MEASURES
The study compared low- and high-disclosure robot personas during a button-initiated collaborative sorting task, measuring participant disclosure and its conversational emergence. Fifty participants completed the procedure after surveys, instructions, and a robot capability demonstration.
- B. Study Conditions: Participants were randomly assigned to low- or high-disclosure conditions, and the robot began small talk about four seconds after the green start-button press with the same opening phrase.The robot always began with: “Hey there, how is your day going so far?”.
- B. Study Conditions: The low-disclosure robot used a basic persona and shared minimal personal information, whereas the high-disclosure robot used an expanded persona with additional personal details.High disclosure was operationalized with plausible, contextually grounded details, including fictional lived experiences or summaries of prior conversations, rather than strongly anthropomorphic or physically inconsistent claims; prior HRI work emphasizes coherent persona and autonomy design,.
- C. Study Procedure: Participants completed consent forms and pre-study surveys, received verbal and written instructions plus a brief demonstration, then sorted PVC fittings by shape and faultiness while the experimenter left.Users indicated the official task start and stop with buttons.
- D. Participants: The study recruited 50 participants, with 25 assigned to each condition; participants’ overall mean age was 23.66 years (SD = 8.29).Participants were compensated $15.00 USD per hour.
- D. Participants: Participants reported mean 1–7-scale experience scores of 5.78 for technology, 4.6 for voice assistants, 3.66 for robots, and 2.36 for teaming with robots.Mean personality self-disclosure was 4.41; the reported standard deviations were 1.20, 1.64, 1.73, 1.52, and 0.78, respectively.
- V. MEASURES: The measures captured both the amount of user self-disclosure and how disclosures emerged during interaction, reflecting that effective small talk involves concise, positive, broad, and contextually relevant responses [23].The study therefore treated conversational dynamics as extending beyond any single metric.
A. Manipulation Check · B. Conversational Measures
The study measured robot speech patterns to verify the intended disclosure personas and quantified user disclosure, response initiation, participation, and utterance types during the main interaction.
- A. Manipulation Check: Robot word count measured the total words spoken during the main task interaction.
- A. Manipulation Check: Robot turns measured the total number of conversational turns during the main interaction.
- A. Manipulation Check: Robot self-disclosure was coded as no, low, or high disclosure, with proportions calculated for each category.Higher disclosure levels represented more personal and elaborated disclosure.
- B. Conversational Measures: User self-disclosure used the same three-level coding scheme, with proportions calculated for each category.The measures captured how personal and elaborated users’ utterances were during the main interaction.
- B. Conversational Measures: Prompted and unprompted responses were recorded as percentages of total turns for users and robots.Independent statements were unprompted, whereas direct responses to preceding statements were prompted.
- B. Conversational Measures: User participation was quantified through total words spoken and total conversational turns during the main interaction.
- B. Conversational Measures: User and robot utterances were classified as questions, imperatives, or non-questions and expressed as percentages of each speaker’s total turns.Non-questions included comments and statements.
C. Subjective Measures (User Perceptions) · VI. DATA ANALYSIS
The study measured users’ perceptions of disclosure, teaming, boredom, and rapport, then analyzed turn-level self-disclosure with mixed-effects models while verifying annotation reliability.
- C. Subjective Measures (User Perceptions): Perceived self-disclosure was assessed with a five-item scale measuring users’ disclosure to the robot during the main task, with Cronbach’s α = 0.80.The construct used a 1–7 response scale.
- C. Subjective Measures (User Perceptions): Perceived teaming was assessed with a five-item scale measuring users’ teaming perceptions, with Cronbach’s α = 0.85.The construct used a 1–7 response scale.
- C. Subjective Measures (User Perceptions): Perceived rapport was measured with the Connection-Coordination Rapport scale, covering rapport-connection and rapport-coordination.Rapport-connection used 12 items with Cronbach’s α = 0.90, while rapport-coordination used 6 items with Cronbach’s α = 0.84, both on a 1–5 scale.
- VI. DATA ANALYSIS: Two human raters independently labeled an initial 10% of the data, reaching Cohen’s κ = 0.909 for utterance type and κ = 0.900 for self-disclosure level after refinement.Study transcriptions were manually verified against video recordings before labeling.
- VI. DATA ANALYSIS: Turn-level user self-disclosure was modeled as an ordinal three-level outcome using a cumulative link mixed-effects model with participant-level random intercepts.The levels were no disclosure, low disclosure, and high disclosure, with higher levels indicating more personal and elaborated disclosure.
- VI. DATA ANALYSIS: The model included experimental condition, prompted versus unprompted user responses, and their interaction as fixed effects.
VII. RESULTS … C. Low Robot Disclosure Leads to High User Disclosure
The manipulation produced distinct robot disclosure behavior without altering conversational speech-act distributions. User disclosure was analyzed as a function of robot disclosure condition and whether responses were prompted.
- A. Manipulation Check: Robot speech differed between conditions: LD produced fewer spoken words than HD, with means of 610.44 and 745.40, respectively.A Student’s t-test found this difference significant, t(48) = −2.125, p = .039.
- A. Manipulation Check: The LD robot produced a significantly higher rate of no-disclosure utterances than the HD robot.
- B. No Condition Differences in Utterance Types: User utterance types did not significantly differ between LD and HD for questions, imperatives, or comments.The reported tests were questions, t(48) = 1.80, p = .08; imperatives, t(30) = 0.40, p = .70; and comments, t(48) = −1.67, p = .10.
- B. No Condition Differences in Utterance Types: Robot utterance types likewise did not significantly differ between conditions for questions, imperatives, or comments.The reported tests were questions, t(48) = −0.74, p = .47; imperatives, t(44) = −1.01, p = .32; and comments, t(48) = 1.31, p = .20.
- B. No Condition Differences in Utterance Types: The disclosure manipulation therefore did not affect the conversation’s structural composition in terms of speech-act distribution.
- C. Low Robot Disclosure Leads to High User Disclosure: User disclosure was modeled at the turn level with a cumulative link mixed-effects model using condition and prompted versus unprompted responses as fixed effects.The ordinal User Self-Disclosure Scores were analyzed with a logit link and participant-level random intercepts.
D. Prior Teaming Experience Shaped Perceived Disclosure, Teaming and Rapport · VIII. DISCUSSION
Participants disclosed more with the low-disclosure robot despite its sharing less, while prior teaming experience shaped perceived disclosure, coordination, connection, and boredom. The findings diverge from prior evidence that agent self-disclosure elicits reciprocity and fosters trust, empathy, and rapport,,.
- D. Prior Teaming Experience Shaped Perceived Disclosure, Teaming and Rapport: LD produced higher disclosure scores than HD (β = 0.94, p = .010), while experienced participants disclosed less with the HD robot (β = −1.84, p = .005).Thus, prior teaming experience amplified the difference in participant disclosure between conditions.
- D. Prior Teaming Experience Shaped Perceived Disclosure, Teaming and Rapport: LD increased perceived teaming (β = 0.93, p = .016) and coordination (β = 0.86, p = .022) relative to HD.Prior robot-teaming experience moderated the coordination effect (β = −0.96, p = .022), with more experienced participants showing reduced HD-condition coordination benefits.
- D. Prior Teaming Experience Shaped Perceived Disclosure, Teaming and Rapport: Condition had no main effects on connection or boredom, but experienced participants reported lower HD-condition connection (β = −0.79, p = .030) and more boredom (β = 1.63, p = .042).These interactions indicate that prior robot experience shaped additional rapport-related outcomes even without overall condition effects.
- VIII. DISCUSSION: The study integrated standardized LD or HD small talk concurrently with physical cooperative work to examine social behavior in realistic, time-limited deployments.The primary systematic difference between conditions was the robot’s disclosure strategy: sharing more or less about itself.
- VIII. DISCUSSION: Participants disclosed more in LD than HD, even though the high-disclosure robot shared more frequently, personally, elaborately, and verbosely.Controlling for covariates, perceived and labeled self-disclosure was significantly lower with the high-disclosure robot.
- VIII. DISCUSSION: The results contradict prior findings that greater dialogue-agent disclosure elicits reciprocal sharing and that disclosure by human-like agents fosters trust, empathy, and rapport,.The discussion identifies embodiment incongruence between the robot’s form and disclosure strategy as one possible explanation for the divergence.
A. Embodiment Incongruence · B. Perceived Authenticity and Limits of Reciprocity
The robot’s non-humanoid embodiment may have made its personal disclosures seem incongruent or inauthentic. Although prompts elicited immediate disclosure, reciprocity remained narrow and did not support broader, genuine engagement.
- A. Embodiment Incongruence: Lower teaming and coordination in the high-disclosure condition suggest that the manipulator’s non-humanoid embodiment may have moderated how personal disclosure was received.The disclosure style may have mismatched participants’ expectations for a conversational partner.
- B. Perceived Authenticity and Limits of Reciprocity: Participants may not have perceived the robot’s disclosures as genuine because its lack of human-like traits conflicted with expectations for a partner that shares personal information.This may explain why the findings opposed expectations from prior literature,.
- B. Perceived Authenticity and Limits of Reciprocity: Interview responses varied, but some high-disclosure participants viewed the robot’s sharing as information about attributes rather than lived experiences.Participants still elaborated on the robot’s disclosure tendencies across conditions.
- B. Perceived Authenticity and Limits of Reciprocity: One participant expected the robot to share first, indicating that the high-disclosure strategy did not reliably satisfy expectations for reciprocal disclosure [H242].This aligns with evidence that recipients disclose more after receiving disclosure first.
- B. Perceived Authenticity and Limits of Reciprocity: Robot prompts elicited disclosure in the moment, but short responses of one or two sentences limited conversational flow and broader engagement.The resulting reciprocity was narrow rather than sustained across the interaction.
- B. Perceived Authenticity and Limits of Reciprocity: Participants questioned the authenticity and depth of the exchange, sometimes ignoring questions because the robot’s responses did not feel genuine.The robot’s prompts could produce immediate answers without creating a meaningful conversational exchange.
- B. Perceived Authenticity and Limits of Reciprocity: Without genuine reciprocity, questions about personal preferences felt superficial and pointless because participants believed the robot lacked corresponding experiences to share [L241].Participants expected reciprocal personal information, such as favorite colors or hobbies, rather than one-sided prompts.
C. Privacy Concerns with Self-Disclosure … 2) Prompted vs Unprompted Responses:
Robot self-disclosure may reduce participant disclosure when it raises privacy concerns, especially among experienced robot collaborators. Prompting elicited immediate responses from high-disclosure participants but did not increase overall conversational engagement.
- C. Privacy Concerns with Self-Disclosure: Sharing the robot’s fictional past conversations may have backfired by prompting less participant self-disclosure through privacy concerns about digital footprints.Participants in both conditions acknowledged potential privacy risks when disclosing to a conversational robot.
- D. Other Factors Shaping Self-Disclosure: Quantitative analyses identified individual differences and prompted responses as additional factors influencing user self-disclosure patterns.These factors supplemented the condition-based explanation of disclosure behavior.
- 1) Prior Robot Teaming Experience:: Prior robot teaming experience shaped disclosure-strategy reception, with experienced participants less receptive to the high-disclosure robot.Stronger mental models of robots may have made persona-based disclosure seem less authentic.
- 1) Prior Robot Teaming Experience:: Participants viewed the robot’s conversational and task-oriented behaviors as “completely two different tasks” [H231], illustrating authenticity concerns.This distinction helps explain why experienced users were less inclined to accept persona-based disclosure as authentic.
- 2) Prompted vs Unprompted Responses:: High-disclosure participants rarely self-disclosed spontaneously, but direct robot prompts elicited disclosure during the interaction.Prompting narrowed the difference between conditions in the moment without increasing overall conversational engagement.
- 2) Prompted vs Unprompted Responses:: Disclosure often clustered around specific conversational turns, consistent with prior communication research rather than being evenly distributed across interaction.The observed prompting effect elicited disclosure in the moment but did not produce broader conversational engagement.
E. Design Implications · IX. LIMITATIONS · X. CONCLUSION
The study recommends calibrating robot self-disclosure to robot form and user experience, while recognizing that findings from a simulated, largely student-based sample may not generalize broadly. Overall, robot disclosure can be double-edged: extensive disclosure reduced reciprocity in this non-humanoid cooperative setting.
- E. Design Implications: On non-anthropomorphic manipulators, favor light, functional, contextually grounded small talk, reserving richer self-disclosure for human-like robots.This design implication matches disclosure richness to the robot’s form.
- E. Design Implications: Use targeted prompts to elicit user disclosure rather than extensive persona-based robot sharing, which may seem inauthentic on functional platforms.The recommendation emphasizes drawing out users instead of over-disclosing through a robot persona.
- E. Design Implications: Adapt robot social behavior to user experience because users experienced in robot teaming are less receptive to persona-based disclosure.Prior robot-teaming experience is identified as a factor that should shape social behavior.
- IX. LIMITATIONS: The study’s simulated quality-control task and largely university-based participant pool may limit generalization to broader populations and real industrial contexts.Real industrial settings may differ in time pressure, task severity, and privacy concerns, while prior robot experience may also affect generalization.
- X. CONCLUSION: Robot disclosure reduced reciprocity in this non-humanoid cooperative setting, with lower user self-disclosure following the high-disclosure robot.The conclusion characterizes disclosure as a double-edged strategy and combines quantitative models with qualitative interviews.
- X. CONCLUSION: The work provides empirical evidence that small-talk disclosure must be carefully calibrated to context and user expectations.The conclusion frames calibration as central to integrating small talk with non-humanoid robots in cooperative settings.