Companion Robots for Older Adults

Project Website GitHub

RoboCom is a line of social robots designed to provide sustained interaction with older adults. The core goal is to induce self-disclosure through a carefully designed conversation agent — a persona of a recently graduated college student who shares stories from their own life to invite the user to share theirs.

XiaoXiao

Data analysis & front-end development · Mar. 2024 – Dec. 2025
International Journal of Human–Computer Interaction, 2026

XiaoXiao is a companion robot for older adults featuring a voice-user interface (VUI) with background storytelling to encourage extended, natural conversation. Primary design and user testing results were presented at ARIS 2024 (Lo et al., 2024).

Key features:

• Background narrative framing to provide conversation context
• VUI enabling fluid spoken interaction without screen dependency
• Iterative user-centered design grounded in gerontechnology

For the experiment comparing a high-affective and a low-affective robot personality (Yueh et al., 2026), I:

• Developed the front-end app that presents the robot’s responses during conversation.
• Defined the robot stimuli semantically for the manipulation check — operationalizing discourse affectiveness into measurable linguistic indicators, and using them to verify that the two robots’ conversational turns actually differed as designed.

Conversations averaged 15 minutes per session and 30 minutes per participant.

Bachelor’s thesis: Zbot, a diary robot

Design and Evaluation of a Conversational Robot for the Elderly Based on Social Penetration Theory

For my bachelor’s thesis I built Zbot, a diary robot for older adults, reusing XiaoXiao’s front end and developing a new back end. Grounded in social penetration theory, the robot is designed to invite older adults to disclose more about themselves over the course of a conversation. The work involved 28 user research sessions with 7 older adults across 4 system iterations.

Measuring self-disclosure from behavior, not questionnaires. Self-report scales of self-disclosure are long, cannot capture behavior as it happens, and are open to dishonest answers. I built a behavioral construct of self-disclosure and a multimodal pipeline to measure it from the conversations themselves, presented at DADH 2025 (Chang & Yueh, 2025):

• Language: LIWC dictionaries to code emotional, cognitive, and social-process markers in what participants said.
• Speech: grounded in Communication Accommodation Theory, vocal entrainment (how speakers converge toward each other) as an index of interpersonal attitude — pitch (F0) and harmonicity with Praat, and spectral features (MFCC, LTAS) with Librosa.

Building Zbot was one of the systems behind the Social Robot Design Platform.