Social Robotics for Disabled Students: An Empirical Investigation of Embodiment, Roles and Interaction

Technical Summary: Social Robotics for Disabled Students

Problem Statement
Students with disabilities in higher education (HE) face significant institutional and social barriers when accessing support, including lengthy procedures, insufficient information, and high social-emotional demands. Existing support mechanisms often fail to meet the needs of these students, with nearly half avoiding raising issues due to procedural complexity and 45% lacking sufficient information on available adjustments. While social robots have shown potential in engagement and assistance, there is a critical gap in understanding how these technologies align with the lived experiences of disabled students. Specifically, it remains unclear how different robot embodiments (physical vs. disembodied) and interaction roles (informational vs. emotional) affect perceptions of understanding, social energy demands, information clarity, task difficulty, and privacy. Furthermore, prior research often relies on medical models of disability or focuses on neurotypical populations, failing to address the heterogeneous nature of disability in human-robot interaction (HRI).

Methodology
The study employed a within-subjects experimental design with 31 disabled students from the University of Cambridge. Participants engaged in a 60-minute session involving six conditions derived from a 3x2 factorial design:

• Roles: Two interaction roles were tested:
  1. Signposting: An informed guide providing actionable guidance and resource direction (e.g., finding lecture recording policies).
  2. Sounding Board: A conversational partner facilitating self-reflection and problem-solving through guided questioning without providing direct solutions.
• Embodiments: Each role was delivered via three modes:
  1. Physical Robot: A humanoid robot (Pepper) utilizing a hybrid dialogue architecture (pre-scripted flows + GPT-4 API responses) with expressive gestures and voice.
  2. Disembodied Agent: A voice-only agent (Amazon Echo Dot) running the identical dialogue logic and speech synthesis as the robot but lacking physical presence.
  3. Control: A standard laptop/web browser interface (for signposting) or a fictive self-referral form (for sounding board) representing the status quo.

Data Collection and Analysis
Data was collected through:

• Task-level Questionnaires: Custom-designed 5-item Likert scales administered after each task, mapping to five research questions (RQs) regarding understanding, social energy, information access, task difficulty, and privacy.
• Pre/Post Questionnaires: Standardized items assessing attitudes toward social robots for disability support.
• Post-study Standardized Scales: The Human-Robot Interaction Evaluation Scale (H-REIS) measuring sociability, agency, animacy, and disturbance; and the Negative Attitudes toward Robots Scale (NARS).
• Semi-structured Interviews: Conducted post-interaction to gather qualitative insights.

Quantitative analysis utilized linear mixed-effects models (accounting for disability type as a fixed effect and participant as a random effect) and non-parametric tests (Wilcoxon signed-rank) due to non-normal data distribution. Qualitative data underwent iterative reflexive thematic analysis.

Key Results

• Perceived Understanding (RQ1): The physical robot was perceived as significantly more understanding than the disembodied agent in both roles. However, this effect was moderated by disability type; the advantage was smaller for participants with Autism, mobility difficulties, and other long-term chronic conditions.
• Social Energy Demands (RQ2):
  • In the signposting task, the control (laptop) condition required significantly less social energy than both the robot and the agent. The robot required more energy than the control.
  • In the sounding board task, no significant differences in social energy were found between the robot, agent, and control.
  • Disability type moderated these effects; for instance, the energy advantage of the control over the agent was smaller for participants with Autism.
• Information Access and Clarity (RQ3): No significant differences were found between the robot, agent, and control regarding information usefulness or clarity.
• Task Difficulty (RQ4): In the sounding board task, the disembodied agent was rated as more difficult than the control. In the signposting task, no significant differences were found between conditions.
• Privacy Concerns (RQ5): In the signposting task, the control was perceived as more private than the disembodied agent, while the robot did not differ significantly from the control. In the sounding board task, privacy perceptions were comparable across all conditions.
• Embodiment and Perception (H-REIS): The physical robot was rated significantly higher on sociability and animacy than the disembodied agent. Conversely, the robot was perceived as more disturbing than the agent overall, though this varied by disability group (e.g., participants with mobility difficulties found the agent more disturbing than the robot, while those with SpLD and Autism found the robot more disturbing).
• Attitudinal Shifts: Post-interaction, participants reported significantly higher apprehension and lower perceived usefulness of social robots compared to pre-interaction scores.

Key Contributions and Significance
The paper claims to be the first sustained body of work investigating how the embodied form of a robot is perceived and engaged with specifically by disabled students in HE. Its primary contributions include:

1. Empirical Evidence on Embodiment: It demonstrates that while physical embodiment enhances perceptions of understanding, sociability, and animacy, it does not universally improve task performance or reduce social energy demands. In some cases (e.g., routine information seeking), a physical robot may impose higher social energy costs than a standard web interface.
2. Role Differentiation: The study highlights that the impact of embodiment is task-dependent. Embodiment offers relational benefits in emotional support contexts but may not offer functional advantages for simple information retrieval.
3. Heterogeneity of Disability: The findings underscore that disability is not a monolith; different disability types (e.g., Autism vs. Mobility difficulties) interact differently with robot embodiment, particularly regarding perceptions of disturbance and social energy.
4. Ethical and Practical Implications: The study identifies that direct interaction with robots can increase apprehension and lower perceived usefulness, suggesting a need for careful expectation management. It argues that for robots to be effective, they must be integrated into institutional infrastructures with robust privacy protections and designed to minimize unnecessary social burdens.

Conclusion
The study concludes that social robots have the potential to mitigate accessibility barriers by providing a non-judgmental, understanding presence, particularly for relational tasks. However, they do not yet outperform self-guided methods for information access and task completion. The authors emphasize that future development must be grounded in the lived experiences of diverse disabled students, moving beyond a "one-size-fits-all" approach to address the complex interplay between embodiment, task type, and individual disability profiles.