Technical Summary: How Does Delegation in Social Interaction Evolve Over Time? Navigation with a Robot for Blind People

Problem Statement

Autonomous navigation remains a significant challenge for individuals with blindness, particularly in complex, dynamic environments such as crowded museums. While robotic systems offer potential for enhancing mobility, fully autonomous systems often fail in dense crowds or when encountering physical obstacles (the "freezing robot problem"). Although shared control—where user input and robot autonomy jointly shape decisions—is recognized as a solution, there is a lack of understanding regarding how blind users perceive and adapt their delegation strategies over time. Prior research has largely relied on short-term trials or Wizard-of-Oz (WOZ) setups where a human operator speaks through the robot, which limits the authenticity of social interactions and fails to capture how trust and interaction strategies evolve with repeated, real-world exposure.

Methodology

The authors conducted a three-week longitudinal study involving six blind or low-vision participants in a real-world museum setting.

System Architecture:
The study utilized a modified open-source, suitcase-style robotic guide equipped with LiDAR and RGB-D cameras. The system featured a shared-control interface with the following capabilities:

• Surrounding GPT: A button-activated feature using GPT to generate broad, open-ended environmental descriptions (objects, people, spatial layout) based on camera input, moving beyond simple safety alerts.
• Obstacle Detection & Contextualization: When the robot stopped, GPT-4o analyzed the camera feed to provide concise, objective explanations of the stop reason (e.g., "There is a crowd ahead with about three people").
• Social Delegation: Buttons allowed users to delegate social interactions, triggering the robot to vocalize specific phrases ("Excuse me, please move" or "Excuse me, please help me") via an external speaker.
• Navigation: The robot navigated autonomously to waypoints, with users holding a handle for physical guidance and issuing commands via a smartphone or handle buttons.

Study Design:

• Participants: Six blind/low-vision individuals (ages 18–69) with varying levels of independence and navigation tool usage.
• Duration: Three weekly sessions, each lasting approximately two hours (including training and navigation).
• Scenarios: Routes were designed to include three specific challenge types:
  1. Crowd Scenarios: Static crowds blocking the path.
  2. Obstacle Scenarios: Physical objects (foam block, suitcases, stroller) blocking the path.
  3. Line Scenarios: Queues at destinations requiring interpretation of social cues.
• Data Collection: Participants completed navigation tasks, followed by post-session interviews and surveys (including the Robotic Social Attributes Scale and study-specific Likert items) to track changes in delegation preferences, confidence, and understanding of the robot's behavior.

Key Contributions

1. Longitudinal Empirical Insights: The study provides rare empirical data on how blind users' perceptions of collaboration and social autonomy evolve over repeated interactions with a fully autonomous robot in a real-world setting, moving beyond single-session novelty effects.
2. Analysis of Evolving Delegation Strategies: It details how participants refined their strategies for delegating social interactions (asking people to move or requesting help) versus acting independently, revealing that preferences are dynamic and context-dependent rather than static.
3. Design Implications for Adaptive Robots: The work offers concrete implications for designing navigation-assistive robots, emphasizing the need for adaptive support levels, interactive description systems that allow for follow-up, and flexible communication styles that accommodate diverse user preferences and social environments.

Results

Evolution of Delegation:

• Shift from Independence to Selective Delegation: Participants generally shifted from exploratory use to purposeful delegation. While some participants (e.g., P2, P3, P4) maintained high delegation rates throughout, others (e.g., P1, P5, P6) initially preferred independence but increasingly delegated social interactions to the robot over time, particularly in noisy or crowded environments where their own voices were less effective.
• Contextual Triggers: Delegation was heavily influenced by situational demands. Users were more likely to delegate when the robot's voice was perceived as carrying better than their own in noisy settings, or when they felt socially anxious about initiating contact.
• Trust Calibration: Participants learned to interpret the robot's micro-movements (e.g., backing up, shifting) as signs of "searching" for a path rather than being stuck. This led to a more calibrated response where users waited for the robot to attempt resolution before intervening, rather than acting on every obstacle alert.

Use of Environmental Descriptions:

• From Curiosity to Strategy: Usage of the "Surrounding GPT" feature evolved from curiosity-driven exploration in Week 1 to targeted, context-specific requests in Weeks 2 and 3. Users began activating descriptions strategically to interpret the robot's behavior (e.g., confirming a crowd density before deciding to delegate) or to gain situational awareness in complex areas.
• Satisfaction: Participants found the GPT-generated descriptions concise and actionable, though some expressed a desire for more interactive, follow-up questioning capabilities.

Confidence and Social Attributes:

• Confidence: While individual trajectories varied, many participants reported increased confidence in acting alongside the robot over time. However, statistical analysis showed that the increase in confidence across the group was not statistically significant, highlighting the influence of situational variability (e.g., crowd density, noise).
• Social Perception: Participants rated the robot positively on competence and reliability. The robot's synthetic voice was often viewed as a tool to overcome social barriers, allowing users to request assistance without the discomfort of direct personal initiation.

Challenges Observed:

• Language Barriers: The robot's verbal requests sometimes failed to elicit responses from non-native speakers, suggesting a need for multilingual support or non-verbal cues.
• Physical Limitations: In cases where the robot became physically stuck (e.g., against slanted walls), participants noted the difficulty of manually moving the heavy device, highlighting a gap in recovery mechanisms.

Significance and Claims

The paper claims that delegation in social interaction is not a fixed preference but a dynamic, adaptive process shaped by repeated exposure, situational demands, and the development of mental models regarding the robot's capabilities.

• Beyond Novelty: The study demonstrates that short-term trials may misrepresent long-term adoption; users' strategies stabilize and evolve only after the novelty effect fades and they gain experience with the system's limitations and strengths.
• Contextual Flexibility: The findings challenge the notion that users universally prefer to take over control when robots struggle. Instead, under certain real-world conditions (e.g., high noise, social anxiety), users may view robot-led interaction as the more effective and comfortable option.
• Design Imperative: The authors argue that future assistive robots must be designed as adaptive partners rather than static tools. Systems should support fluid negotiation of control, offering transparency in robot state (e.g., distinguishing between "searching" and "stuck"), providing adaptable levels of descriptive detail, and accommodating diverse user personalities (from "independent-first" to "delegation-first" users).

The paper concludes that for navigation-assistive robots to achieve long-term adoption, they must be evaluated in longitudinal settings to understand how trust and collaboration evolve, ensuring systems are built to support independence while remaining responsive to the complexities of real-world social navigation.