PaperlayerView exact revision on arXiv

Source-linked AI summary

A Deployable Architecture for Robot-Mediated Tasks (DART): Evaluation in Socially Assistive Robot-Guided Cognitive Behavioral Therapy Exercises

Mina Kian, Lydia Ignatova, Jiong Wang, Ji Min Lee, Jiancheng Li, Qianwei Guo, Emily Weiss, Amy O'Connell, Kaitlin Zareno, Jiani Li, Reyna Patel, Leyaa George, Minyu Huang, Justin Yang, Maja J. Matarić

arXiv:2609.12349v1cs.RO

TL;DR

Low-cost SARs have limited onboard capabilities for the complex interactions and data needs of longitudinal real-world use. DART distributes those capabilities across a robot, web application, and cloud infrastructure; its CBT system evaluations reported usability and technical feasibility, while not testing clinical efficacy.

  • Problem

    Low-cost SARs have limited onboard capabilities for visual interaction, user input, data storage, remote computation, and other demands of real-world deployment.

    Source: §I · p. 1, C. Low-Cost and Open-Source SAR… · p. 3

  • Method

    DART distributes interaction between a low-cost robot, a web application, and cloud infrastructure, and was instantiated in a CBT homework system through iterative design and user studies.

    Source: §I · pp. 1–2

  • Results

    The evaluations established the system’s usability and technical feasibility, with participants describing it as enjoyable and easy to use.

    Source: 3) Participant Feedback · p. 12, §IX · pp. 16–17

  • Takeaways & Limitations

    DART provides a flexible foundation for accessible, longitudinal SAR systems with distinct roles for the robot, web application, and cloud infrastructure.

    Source: §IX · p. 17

    The in-lab study lacked a control condition, and the in-home study had a small, demographically homogeneous sample; the evaluations did not aim to demonstrate clinical efficacy.

    Source: §IX · p. 16

Abstract

from arXiv · show

Socially assistive robots (SARs) can support structured health and well-being interventions, but hardware and cost constraints limit interaction complexity and longitudinal real-world deployments. We present DART: Deployable Architecture for Robot-Mediated Tasks, an architecture that extends SARs through a web application and cloud infrastructure, enabling visual content, user input, remote computation, and persistent data storage synergistically with the robot's physical embodiment, speech, and movement. We evaluated DART by instantiating it in an interatively-developed full-stack HRI system for helping university students with elevated generalized anxiety to complete cognitive behavioral therapy (CBT) homework exercises. The resulting system, which used the low-cost open-source Blossom robot platform, was refined and evaluated through a participatory design process and multiple user studies, and finally evaluated in an in-lab study with 103 participants, and then a six-week in-home deployment with four participants. In the in-lab evaluation, participants showed significant within-session reductions in stress, state anxiety, and negative affect, and gave the platform a mean System Usability Scale score of 78.89. In the home deployment, the mean System Usability Scale score was 87.5, with positive qualitative feedback on usability. Participants across both groups identified speech input, visual presentation, and web-robot synchronization as priorities for improvement. These findings validate DART as an effective architecture for extending the capabilities of a low-cost SAR in both in-lab single-session and in real-world longitudinal deployments.

I. INTRODUCTION

Low-cost SARs can support rich, sustained interactions, but limited onboard capabilities and the demands of real-world deployment create technical and logistical challenges. DART distributes capabilities across a robot, web application, and cloud infrastructure.

I. INTRODUCTION

  • DART extends low-cost robots by assigning physical interaction to the robot and visual content, user input, and remote services to the web and cloud.This division addresses capabilities that are often impractical to incorporate onboard.Source: §I · p. 1

A. Socially Assistive Robots (SARs)

  • Embodied robots have been associated with stronger social presence, engagement, and task performance than virtual agents, and can support motivation during sustained interventions.Studies also report greater social responses and compliance toward physically present robots than video-displayed versions.Source: A. Socially Assistive Robots (SARs) · p. 2

B. Real-World and Longitudinal SAR Deployments

  • Longitudinal SAR deployments reveal how users integrate robots into daily life, but require reliability, data continuity, privacy, affordability, and maintenance with limited researcher supervision.Engagement may also change as novelty diminishes.Source: B. Real-World and Longitudinal SAR… · p. 2

C. Low-Cost and Open-Source SAR Platforms

  • Low-cost open-source platforms offer customization and lower fabrication and maintenance costs, but constrained onboard computers may not support complex interfaces or intensive processing.Adding more capable hardware can increase cost, power consumption, and maintenance complexity.Source: C. Low-Cost and Open-Source SAR… · pp. 2–3

D. Extending Robots Through External Interfaces

External interfaces and cloud infrastructure can extend robots without substantial changes to their embodiment, while distributing interaction, computation, and data management across available devices. DART’s design goals include accessible, modular, synchronized, and secure longitudinal use.

D. Extending Robots Through External Interfaces

  • Web interfaces can leverage users’ existing devices for visual content, detailed responses, surveys, and accessibility features without adding dedicated robot components.Dedicated tablets or stations otherwise add hardware to purchase, configure, transport, and maintain.Source: D. Extending Robots Through External… · p. 3

E. Cloud and Edge Architectures for Human-Robot Interaction

  • Distributed HRI systems assign embodied, time-sensitive behavior to robots and use local devices and cloud infrastructure for content, coordination, computation, and persistent storage.This modularity supports scalable deployments but brings network reliability, synchronization, privacy, and security challenges.Source: E. Cloud and Edge Architectures for… · p. 3

A. Design Goals & Requirements

  • DART’s goals include extending low-cost robots, supporting independent longitudinal use and modular structured activities, synchronizing web and robot behavior, and protecting sensitive data.The architecture requirements include authenticated remote connections, controlled access, and encrypted transmission and storage.Source: A. Design Goals & Requirements · pp. 3–4

B. Architecture Components

DART separates interaction orchestration and data management from robot-specific behavior execution across a web application, cloud services, and a network-connected robot. The components coordinate through cloud messaging while leaving latency-sensitive physical control local.

1) Robot:

  • The robot executes high-level behavior commands and returns status or error messages, while local software handles latency-sensitive operations and translates commands into platform-specific actions.An Internet-connected controller and MQTT support integration without direct web-application access to robot hardware.Source: 1) Robot · p. 4

2) Web Application:

  • A browser-based application presents instructions and interactive content, collects input, and coordinates robot behavior with activity and session progression.It can use a device’s camera and microphone, subject to browser permissions, for configured audiovisual collection.Source: 2) Web Application · p. 4, 1) Robot · p. 4

3) Cloud Services & Data Storage:

  • Cloud services provide communication, authentication, persistent session state, data storage, and remote computation for the web application and robot.Structured responses and progress are stored separately from larger audiovisual files.Source: 3) Cloud Services & Data Storage · p. 5

4) Communication and Synchronization:

  • DART uses a cloud-hosted MQTT publish–subscribe model in which the web application sends high-level commands and the robot publishes status, completion, or error responses.The application uses robot responses to synchronize on-screen content and determine when an interaction can progress.Source: 4) Communication and Synchronization · p. 5
  • Encrypted outbound connections to the cloud broker allow the web application and robot to operate across different local networks without inbound connections or firewall changes.The web application uses MQTT over secure WebSockets, while the robot uses MQTT-over-TLS.Source: 4) Communication and Synchronization · p. 5

5) Security and Privacy:

DART combines safeguards for data and access with a modular activity-development model. Its CBT homework instantiation illustrates requirements for structured, independently usable activities that coordinate robot behavior and web content.

5) Security and Privacy:

  • DART encrypts communications and stored data, restricts access through authentication and role-based permissions, and limits sensitive data stored on edge devices.Camera and microphone access requires browser permission and recording is restricted to activities that need audiovisual data.Source: 5) Security and Privacy · p. 5

C. Application Development Model

  • Activities issue reusable high-level robot commands, while robot-specific handlers translate them into platform behaviors, separating activity logic from hardware implementation.Hosted activity modules can be revised or added without modifying deployed robots or requiring user software installation.Source: C. Application Development Model · p. 5

A. Evaluation Context

  • The DART instantiation delivers CBT homework exercises for university students with elevated general anxiety, a context requiring structured instructions, reflection, visual information, prompts, and guided practice.The six-week deployment was designed for independent home use, requiring persistent progress, authenticated access, and operation on home networks.Source: §IV · p. 5, A. Evaluation Context · p. 6
  • During exercises, the website presented instructions, visuals, and response fields while the robot provided spoken guidance and movement at designated points.The platform also administered surveys and stored participants’ responses.Source: A. Evaluation Context · p. 6

B. Evaluation System

DART pairs Blossom’s embodied behaviors with a browser-based interface and cloud services. This distribution supports structured activities, sensing, remote computation, data storage, and synchronization.

1) Robot Design:

  • Blossom is a roughly US $400 open-source tabletop robot whose Raspberry Pi locally executes speech, movement, and other high-level commands.Reusable behaviors include speech, ear-color changes, backchannel movements, and predefined activity sequences.Source: 1) Robot Design · p. 6

2) Web Platform and Language-Model Integration:

  • The laptop-based web platform presents instructions and media, collects responses, manages CBT activity progression, and coordinates screen content with Blossom’s speech and movement.Researchers could update hosted activity modules during deployment without requiring participants to install software or return the robot to the lab.Source: 2) Web Platform and Language-Model… · pp. 6–7
  • For generated dialogue, the web and cloud components call GPT-4o, allowing responses to be delivered through the robot without running the model on its onboard computer.The browser can also record audiovisual data, run lightweight face detection locally, and transfer recordings to cloud storage.Source: 2) Web Platform and Language-Model… · p. 7

3) Cloud Services and Data Storage:

  • AWS hosts the platform, stores participant responses and recordings, brokers web–robot MQTT messages, and supplies remote computation such as synthesized speech.Transit and stored data are encrypted, access is restricted, and robot credentials are limited to required operations and topics.Source: 3) Cloud Services and Data Storage · p. 7

4) In-Home Deployment and Connectivity:

The system was prepared for home connectivity and refined through participatory design. Student and therapist feedback informed system changes and highlighted priorities for a coherent, customizable experience.

4) In-Home Deployment and Connectivity:

  • Researchers preconfigured home Wi-Fi so the robot connected to AWS IoT Core at startup, enabling communication across local networks without firewall changes.This setup let participants begin without configuring the robot and avoided researcher home visits for network setup.Source: 4) In-Home Deployment and Connectivity · p. 7

V. PARTICIPATORY DESIGN PROCESS

  • Participatory design examined whether the web application extended the robot, supported independent use over time, and worked with the robot as a coherent multimodal system.Feedback on the robot, interface, activities, and their integration informed refinements before subsequent deployment.Source: §V · p. 7

A. Data Collection

  • The design study involved 15 students with elevated self-reported generalized anxiety and 15 therapists, using demonstrations, structured activities, group discussions, and recorded sessions.The analysis in this paper focused on participants’ feedback about the robot and web application implementation.Source: A. Data Collection · pp. 7–8

B. Data Analysis

  • Researchers used hybrid thematic analysis to cluster student and therapist transcript observations, and separately organized worksheet responses to compare verbal and written feedback.The separate worksheet analysis retained ideas that appeared only in written responses.Source: B. Data Analysis · p. 8

1) Key Design Priorities:

  • Four themes emerged: embodiment and expression, speech and conversation, interface and input, and sensing and adaptation; participants also wanted customizable behaviors and clearer guidance.Preferences for expressive movement and physical interaction differed, with some finding these features distracting or wanting more control.Source: 1) Key Design Priorities · p. 8

2) Design Changes:

Participatory-design feedback prompted system refinements, while some requested features remained deferred. The in-lab evaluation then examined CBT exercise effects and the system’s usability, interaction, and technical performance.

2) Design Changes:

  • Design changes added controls for movement expressiveness and ear color, reduced distracting swivel movements, and introduced touch-responsive behavior.Speech controls and shorter, more natural language-model responses were added while retaining CBT exercise structure; dashboard and support-resource changes also clarified the interface.Source: 2) Design Changes · p. 8

3) Later and Deferred Features:

  • A mobile web version and more extensive affective sensing, heating, and weighted features were deferred; mobile access was not implemented because reliable webcam recording was required.Fitbit data collection was included in later evaluations, but physiological data did not directly alter robot behavior in real time.Source: 3) Later and Deferred Features · p. 8, 2) Design Changes · p. 8

VI. IN-LAB EVALUATION

  • The in-lab evaluation assessed whether CBT modules were clinically promising, whether exercises changed within-session state anxiety, and whether web content and robot behavior stayed synchronized.The evaluation followed participatory-design-informed improvements and also examined the system’s visual-content, input, and data-collection capabilities.Source: §VI · p. 8

A. Participants & Recruitment

  • The in-lab study enrolled 103 participants with GAD-7 scores of at least 5, excluding people whose screening indicated moderate-to-high suicide risk.Participants were ages 18–49, with a mean age of 22.07.Source: A. Participants & Recruitment · pp. 9–10

B. Study Design and Procedure

  • Participants completed three exercises in an approximately 90-minute individual session, with chatbot topics, additional exercises, and exercise order counterbalanced across conditions.They completed baseline and post-exercise psychological measures, while the system recorded responses and session video and collected Fitbit heart-rate data.Source: B. Study Design and Procedure · p. 10

C. Measures

The evaluation measured perceived usability and immediate psychological state at repeated points during a session.

1) System Usability:

  • The 10-item System Usability Scale was administered after all exercises and scored from 0 to 100, with higher scores indicating greater perceived usability.Participants could optionally provide a free-response comment.Source: 1) System Usability · p. 11

2) Psychological State:

  • Stress, state anxiety, and positive and negative affect were measured before the first exercise and after each exercise, yielding four observations per participant.Participants could decline individual survey questions, so the number of observations varied slightly across analyses.Source: 2) Psychological State · p. 11, C. Measures · p. 11
  • Stress used an 11-point current-stress rating, while anxiety used adapted state-anxiety items and affect used separate positive and negative PANAS-SF scores.Subsequent anxiety assessments used three items to reduce repeated-measure burden; higher scores indicated greater stress and anxiety.Source: 2) Psychological State · p. 11

3) Post-Session Exit Interviews:

The in-lab evaluation found significant within-session improvements across four self-reported psychological measures, alongside good perceived usability. Exit interviews and written comments were reviewed descriptively to identify system feedback.

3) Post-Session Exit Interviews:

2) Psychological State:

  • Psychological-state measures were analyzed separately across complete four-timepoint cases using repeated-measures ANOVA and Holm-adjusted pairwise tests.Greenhouse-Geisser corrections were applied when sphericity was violated.Source: 2) Psychological State · p. 11

1) System Usability:

2) Psychological State Changes:

  • Stress and negative affect decreased after each exercise, while anxiety and positive affect did not change significantly after the third exercise.The overall baseline-to-session-end decreases in stress and anxiety and increase in positive affect were significant; the overall negative-affect decrease was also significant.Source: 2) Psychological State Changes · p. 12
  • Stress, state anxiety, positive affect, and negative affect all improved significantly within the session, with stress and negative affect changing after each exercise.Anxiety and positive affect improved mainly during the first two exercises and then remained relatively stable; without a no-treatment control, changes cannot be attributed exclusively to CBT exercises.Source: 2) Psychological State Changes · p. 12, Figure 6 · p. 12

3) Participant Feedback:

In-lab feedback was generally positive, while participants also identified specific interaction and interface improvements. The in-home evaluation examined independent use of the system over six weeks with four participants experiencing elevated anxiety.

3) Participant Feedback:

  • In-lab feedback came from exit interviews with all 103 participants and optional written comments from 38 participants.The descriptive review focused on ease of use, input preferences, visual design, and website–robot coordination.Source: 3) Participant Feedback · p. 12
  • Participants wanted the option to speak responses to Blossom, which they felt could make interaction more natural and less like using a separate computer interface.One optional comment also suggested speech input could improve accessibility for older adults and people with disabilities.Source: 3) Participant Feedback · p. 12
  • Website opinions varied: many interviewees found it clear and easy to navigate, while a smaller group wanted more color, polish, or interactive elements.Approximately 23% discussed the interface in interviews; about 18% of the full sample described it positively and about 6% suggested changes.Source: 3) Participant Feedback · p. 13
  • Participants generally described the system as enjoyable and easy to use, while identifying speech input, visual presentation, and more reliable web–robot synchronization as priorities.A handful reported that the website sometimes still indicated Blossom was speaking after it had finished, requiring manual advancement and creating uncertainty.Source: 3) Participant Feedback · pp. 12–13

VII. IN-HOME EVALUATION

  • The six-week in-home pilot assessed independent use, remote maintenance, and transmission and storage of participant responses and activity records.Four participants completed the evaluation; all had elevated anxiety based on the study’s screening criteria.Source: §VII · p. 13, A. Participants and Recruitment · p. 13

B. Study Design and Procedure

The in-home study deployed the web-and-robot system for six weeks, with four CBT exercises requested each week. Participants rated usability highly but also described technical disruptions and preferences spanning interaction and activity formats.

B. Study Design and Procedure

  • Onboarding covered system setup, robot connectivity, exercise use, Fitbit care, troubleshooting, and contacting the research team; participants took the robot, Fitbit, and chargers home.The equipment loaned is shown in Figure 7.Source: B. Study Design and Procedure · p. 13
  • Participants were asked to complete four CBT exercises each week for six weeks, using the website for instructions and input and Blossom for corresponding speech and movement.Exercises included both physiological and text-based activities.Source: B. Study Design and Procedure · pp. 13–14
  • The deployment recorded surveys, exercise inputs, robot power events, settings changes, and session videos, storing data in Amazon S3 and DynamoDB.These records were used to assess whether interactions and participant responses were captured during the deployment.Source: B. Study Design and Procedure · p. 14

C. Measures

  • Participants completed the SUS two weeks into deployment, and one researcher used hybrid inductive–deductive thematic analysis on exit interviews from all four participants.SUS scores range from 0 to 100; interview analysis identified recurring themes and suggested system improvements.Source: C. Measures · p. 14, D. Data Analyses · p. 14

E. Results

  • Three of four in-home participants completed the SUS, with a mean score of 87.5 and all three scores in the “excellent” range.Scores were 85.0, 87.5, and 90.0; the study examined them descriptively given the small sample.Source: E. Results · p. 14
  • All four participants encountered intermittent technical problems, including unsaved data, incorrect activity status, robot connection or speech interruptions, and recording-upload failures.Some failures forced participants to restart activities or repeat responses; recommendations included clearer progress, charging, and connection guidance.Source: E. Results · p. 14
  • Participants responded positively to Blossom’s appearance, and some associated its physical presence and spoken interaction with companionship or feeling listened to.The crochet exterior, ears, and customizable lights were described as approachable and endearing.Source: E. Results · p. 14
  • Participants differed in preferences for structured versus open-ended activities, and some wanted more physical engagement or a better balance between formats.They also wanted additional unstructured conversations; the modular curriculum allowed activity content and ordering to be adapted during deployment.Source: E. Results · pp. 14–15

VIII. DISCUSSION

The discussion presents DART as a way to distribute interaction capabilities across a low-cost robot, web application, and cloud services. Study findings also point to design needs for coherent multimodal interaction, adaptable activities, and longitudinal support.

A. Architecture-Level Findings

  • DART assigns physical presence, speech, and movement to Blossom, visual content and input to the web application, and authentication, communication, computation, and storage to cloud services.This arrangement supported more extensive visual and computational capabilities while keeping the robot comparatively simple and affordable.Source: A. Architecture-Level Findings · p. 15
  • In-lab and in-home mean SUS scores were 78.89 and 87.5, respectively, with participants describing the system as simple, intuitive, or easy to use.The in-home pilot focused on successful deployment, data collection, and ease of use with limited difficulty.Source: A. Architecture-Level Findings · p. 15
  • Participants’ technical difficulties prompted requests for clearer setup instructions, charging reminders, and visual guidance for power cycling the robot.The discussion identifies maintenance reminders and user-performable recovery procedures as important when researcher assistance is limited.Source: A. Architecture-Level Findings · p. 15
  • DART depends on users’ personal devices, internet connectivity, and cloud services, potentially creating access barriers or interruptions and shifting technical demands to participants.The discussion prioritizes remote diagnostics and recovery for longitudinal settings where researchers are not physically present.Source: A. Architecture-Level Findings · p. 15
  • Researchers could update hosted activity content and sequencing after deployment while reusing robot behaviors, reducing the need to modify the robot for each content change.Additional devices and data sources, such as a Fitbit, could be incorporated into specific instantiations without substantially modifying the broader architecture.Source: A. Architecture-Level Findings · p. 15
  • Participants’ varied preferences for activities support adapting modular curriculum content and ordering rather than relying on a single activity format.The in-home preferences spanned psychoeducation, cognitive exercises, grounding, breathing, and open-ended conversation.Source: A. Architecture-Level Findings · p. 15
  • Text and speech serve different preferences: speech may offer conversational immediacy, while text may better support privacy, precision, and reflection.The system retained text input while adding speech input in a later version.Source: A. Architecture-Level Findings · p. 15
  • Synchronization mismatches disrupted the perceived coherence of the robot and website, motivating improved coordination and recovery behavior.The discussion calls for explicit completion signals, timeouts, and robust recovery protocols.Source: A. Architecture-Level Findings · p. 16

B. Findings from the CBT Use Case

In-lab participants showed short-term psychological changes, and in-home participants responded positively to the CBT modules. The authors treat these findings as preliminary because the evaluations do not establish clinical efficacy or isolate the effects of the CBT content, robot, and interface.

B. Findings from the CBT Use Case

  • In-home participants enjoyed the CBT modules and reported learning the curriculum skills; familiar participants said the modules matched their expectations.A participant less familiar with CBT expressed interest in pursuing therapy after the interaction.Source: B. Findings from the CBT Use Case · p. 16
  • Stress and negative affect improved significantly after each exercise; state anxiety and positive affect improved mainly during the first two.The reported changes were within-session.Source: B. Findings from the CBT Use Case · p. 16
  • Without no-intervention or website-only controls, the changes cannot be attributed specifically to CBT exercises, robot embodiment, or the integrated system.Repeated measurement, time, and increased comfort in the study environment are also possible explanations.Source: B. Findings from the CBT Use Case · p. 16

IX. CONCLUSION

  • DART coordinated a robot, web application, and cloud infrastructure in one multimodal interaction, with in-lab and in-home studies demonstrating usability and identifying improvement priorities.The in-lab evaluation involved 103 participants; the in-home evaluation lasted six weeks with four participants.Source: §IX · p. 16
  • The evaluations were limited by the in-lab study’s absent control condition and the in-home study’s small, demographically homogeneous sample.The work was not intended to establish clinical efficacy or isolate effects of CBT content, embodiment, and the browser interface.Source: §IX · p. 16
  • DART provides a flexible foundation for longitudinal SAR systems with defined component roles, reliable coordination, and recovery when components fail.Source: §IX · p. 17
Loading 2609.12349v1…