Robot-Assisted Decision-Making: Unveiling the Role of Uncertainty Visualisation and Embodiment

Title of the Paper

Robot-Assisted Decision-Making: Unveiling the Role of Uncertainty Visualisation and Embodiment

Paper Information

• Domain: Human-Robot Collaboration, Visualization Decision Support, Human-Robot Interaction (HRI)
• Keywords: Decision Support, Artificial Intelligence, Uncertainty Visualization, Robot Embodiment, Risk Communication, Transparency, Trust, Data Visualization

Research Background and Problems

• Problems and Challenges:

  • In high-risk collaborative tasks, robots need to communicate their operational uncertainty to humans effectively. However, there is currently a lack of information on how robots can visualize and convey uncertainty effectively.
  • Although the fields of Human-Robot Interaction (HRI) and Information Visualization (VIS) have studied "uncertainty" and "decision support," there is limited research on the specific application of robot uncertainty visualization in decision-making tasks.
  • There is insufficient understanding of how robots presenting high/low confidence levels impact human decision-making, especially in collaborative scenarios involving physical presence.

• Significance:

  • Enabling robots to convey uncertainty can enhance transparency, calibrate human trust in machines, and support better collaborative decision-making.
  • In critical scenarios (e.g., medical or industrial high-risk tasks), incorrect decisions may pose threats to human safety.

• Motivation and Related Work:

  • Current robots primarily focus on recommendation tasks in social or service contexts, with limited research examining robots in non-humanoid, non-social environments visually conveying uncertainty.
  • In HRI, the role of physically present robots (e.g., robotic arms) in collaborating with humans and conveying uncertainty through behavior warrants further exploration.

Proposed Solution

• Proposed Method or Solution:

  • Investigate two visualization methods (Graphical User Interface (GUI) and embodied behavior) for conveying robot uncertainty and their impact on user decision-making.
  • Conduct experiments using a robotic arm to evaluate its effectiveness in conveying decision information with varying confidence levels (low, high, 100%) during collaborative tasks.

• Innovations:

  • Compare traditional visualization techniques (e.g., icon arrays, dashboards, bar charts) with robot embodied behaviors (e.g., hesitant gestures).
  • Test the feasibility of using embodied behaviors instead of screen-based displays to convey uncertainty in real-world scenarios (e.g., medication packaging and boxing).

• Implementation Steps and Techniques:

  1. Research Questions:
     • RQ1: How do different uncertainty presentation methods affect user decision-making behavior?
     • RQ2: How do varying confidence levels impact users' trust and perception of robot transparency?
  2. Experimental Design:
     • A total of 36 participants were randomly assigned to either the GUI or embodied behavior experimental group to perform a medication boxing task.
     • Each group was tested under three confidence conditions (high, medium, low), with 10 observations and decisions per condition.
  3. Experimental Task:
     • The robot presented confidence levels and coordinated human-robot tasks, requiring participants to decide whether to box and transport medication or conduct further testing based on the robot's suggestions.
  4. Measurement Metrics:
     • Decision accuracy, average confidence estimation, trust and transparency perception during the task.

Research Findings

• Specific Results:

  • Decision Behavior:
    • Under the 100% confidence condition, participants using embodied behavior had a higher correct decision rate (12/18 participants) compared to the GUI group (7/18 participants).
    • Uncertainty visualization methods significantly influenced decision behavior, with embodied behavior encouraging more intuitive decisions, while GUI helped participants evaluate risks more thoroughly.
  • Information Communication and Trust:
    • Users' perception of robot transparency significantly decreased under low confidence conditions, indicating a need to improve explanations of low-probability risks.
    • GUI provided detailed and reliable information for high-risk tasks, but its high data "salience" could lead to "overthinking," resulting in excessive caution.
    • Under the 100% confidence condition, GUI group participants tended to perceive the robot as "overconfident," raising trust issues in human-robot interaction.
  • Advantages of Embodied Behavior:
    • Intuitive presentation requiring no additional learning; applicable to robots with limited information display capabilities (e.g., robotic arms).

• Comparative Advantages Over Existing Solutions:

  • Integrates methods from VIS and HRI fields to study the impact of information presentation on machine-human collaboration, trust, and interaction transparency.
  • Highlights the trade-offs between two decision-making approaches: GUI is better suited for complex decisions but prone to overload, while embodied behavior is more intuitive but conveys less information.

• Experimental or Evaluation Results:

  • Mixed experimental design under different conditions revealed significant impacts of confidence levels, presentation modalities, and trust and transparency perceptions.
  • Transparency scale scores showed no significant difference between embodied behavior and GUI groups, indicating that embodied behavior can partially replace traditional graphical interfaces.

• Limitations and Future Directions:

  • Did not cover more complex decision categories (e.g., multi-objective trade-offs).
  • Further exploration of embodied behavior's applicability in dynamic error detection scenarios is needed.
  • Suggest studying how to dynamically adapt robots to users' varying confidence level needs and designing "behaviorally consistent" robot uncertainty visualization schemes that are less prone to misuse.

Conclusion

This study reveals that uncertainty visualization significantly impacts decision-making and trust issues in human-robot interaction during collaborative tasks. The authors propose further research into the adaptation of uncertainty presentation methods to task types and complexities, particularly in critical high-risk scenarios, to optimize the use of robot embodied behaviors and traditional GUI-based uncertainty visualization methods.