Investigating LLM-Driven Curiosity in Human-Robot Interaction

Research Background and Problem

• What problems or challenges did the authors identify?
  Embedding curiosity behaviors in robots can enhance their learning and adaptability, but the impact of robots' "curiosity" on user perception, interaction, and experience remains unclear. Additionally, most existing studies focus on computational curiosity in robot learning, with a lack of research on "user-centered curiosity behaviors" (i.e., curiosity expression behaviors centered on user experience). Another significant challenge lies in balancing task-oriented outcomes with exploratory behaviors of robots.

• Why is this problem important?
  Integrating curiosity behaviors into robotic systems can simulate more human-like and autonomous robot behaviors, thereby enhancing user experience and the naturalness and enjoyment of human-robot collaboration. Furthermore, understanding how users perceive curious robots is crucial for future robot design and user experience optimization.

• Research Motivation and Related Work
  Computational curiosity is primarily used to address the self-exploration of unseen data in artificial intelligence and robotics. However, most studies on curiosity behaviors focus on optimizing learning efficiency, such as user annotation interactions in active learning, with limited attention to behavior perception in real-world interaction scenarios. Moreover, traditional rule-based robotic systems struggle to model complex social characteristics, while the development of large language models (LLMs) provides an effective solution for dynamically and flexibly generating behaviors aligned with specific personality traits.

Solution

• What methods or solutions did the authors propose?
  The authors proposed a robot curiosity behavior generation framework based on multimodal large language models (MM-LLM). By designing system prompts based on role descriptions, the authors embedded "curious" or "non-curious" personality traits into the robot to drive its expression of non-verbal and verbal curiosity behaviors.

• What are the innovative aspects of this solution?

  1. Utilizing MM-LLM to dynamically generate multimodal curiosity behaviors, integrating vision, language, and physical manipulation to achieve dual goals of social and task orientation.
  2. Implementing curiosity behaviors through role descriptions in natural language, with dynamic behavior strategy adjustments based on user interaction.
  3. Applying curiosity behaviors to real collaborative tasks and conducting cross-task validation to test their generalizability and user perception.

• What are the implementation steps and key technologies used?

  1. System Design:
     • The authors designed a multimodal robotic system with capabilities including visual perception, manipulation actions (e.g., shaking, probing), language generation, and speech recognition.
     • Two role personalities (curious and non-curious) were defined through natural language to establish role behavior rules and priorities.
  2. Curiosity Behavior Implementation:
     • Social curiosity: Asking users about their names and preferences.
     • Cognitive exploration: Asking task-related questions to reduce cognitive uncertainty.
     • Perceptual exploration: Actively exploring the environment using physical actions (e.g., shaking or observing objects).
  3. User Study:
     • Testing the robot under curious and non-curious conditions with 20 participants collaborating on pizza-making or cocktail-mixing tasks.
     • Analyzing user behavior and perception through questionnaires, interaction logs, and interviews.

Research Findings

• What specific findings were achieved?

  1. Users generally perceived the system's exhibited curiosity behaviors: the "curious" system demonstrated significantly higher levels of social curiosity, exploratory curiosity, and task-related curiosity.
  2. Participants preferred collaborating with curious robots. 70% of participants expressed a preference for the "curious" robot, describing it as more interactive, human-like, and engaging.

• What advantages does it have compared to existing solutions?

  1. User Experience: Curiosity behaviors significantly increased the robot's anthropomorphism and animacy, enhancing the naturalness of interactions and user satisfaction.
  2. Generality: The system demonstrated high consistency and robustness across two different scenarios (pizza-making and cocktail-mixing tasks), proving the generalizability of the personalized design.
  3. Technical Flexibility: By adjusting system prompts, role behavior can be quickly switched (e.g., from "curious" to "non-curious"), reducing system development complexity.

• What were the experimental or evaluation results?

  • Quantitative Analysis: The curious system exhibited significantly higher interaction turn density and user-system switching frequency, indicating that curiosity behaviors better stimulated user interaction.
  • Questionnaire Results: Under the "curious" condition, users gave significantly higher ratings for the system's social curiosity attributes (PSC scale), anthropomorphism, and animacy (Godspeed scale).
  • User Feedback: Participants commonly noted that curiosity behaviors enhanced the system's proactivity, interactivity, and learning ability, particularly through questioning and physical exploration.

• Limitations and Future Directions

  1. Short-term Observation Limitation: The experiments were conducted in single-task, short-term scenarios, lacking observations of long-term use and dynamic learning. Future research could explore the performance and user perception of curious robots in continuous learning contexts.
  2. Task Complexity: Current tasks were simple, and some users felt that the simplicity limited opportunities for curiosity behaviors to manifest. Future applications should test potential in more complex tasks and environments.
  3. System Customization: Currently, behaviors are formed through explicit rules. Future research could explore reducing explicit prompts to enhance generalizability.

Conclusion

This study successfully developed and validated a curiosity-driven robotic system based on MM-LLM, significantly enhancing sociality, interactivity, and user satisfaction in interactive contexts. This advancement provides inspiration for future robot design, demonstrating that large language models can effectively shape and dynamically adjust system personality traits and behavioral expressions. In the future, curiosity behaviors are expected to become a critical feature for balancing task efficiency and user experience in social robots.