The Droodle Task: Propelling Human Creativity with LLMs
This project, conducted at the Robot Studio in the Department of Robotics, University of Michigan, explores how large language models (LLMs) can act as collaborative partners in enhancing human creativity through the Droodle Captioning Task. In collaboration with Patricia Alves-Oliveira and Peter H. Kahn, Jr., and robotics student Trey Davis, this research leverages AI-driven assistance to stimulate novel, creative responses from human participants in a structured, engaging way. Ultimately, the Droodle framework will be embedded into a robot, creating an interactive and embodied AI-driven creative experience.
Project Overview:
The “Droodle Captioning Task” presents users with abstract hand-drawn images (droodles), prompting them to generate surprising captions that shift the viewer’s perception of the drawing. This study examines how LLMs can foster creativity by guiding users through a structured process that combines visual and verbal elements.
Current Development:
Framework Design & Implementation: The framework has been designed and is in active development, with the initial stages nearing completion.
User Interface (UI): A custom UI has been created to facilitate user studies, enabling participants to interact fluidly with the system while generating creative captions.
Advancement toward Multi-Agent Models: The next phase includes implementing a multi-agent setup, where agents will challenge each other’s ideas, encouraging richer and more diverse outputs.
Key Innovations:
Creative Collaboration Modules: Structured AI interaction frameworks to balance divergent (idea generation) and convergent (refinement) thinking.
Process-Oriented Prompting: Developed prompts to help LLMs foster creative interactions, sustain user motivation, and avoid overwhelming options.
Multi-Agent Exploration: Future developments will involve a model where multiple agents interact and challenge each other’s outputs, further stimulating creative diversity.
Research Areas:
Computational Creativity
Human-AI Collaboration
Natural Language Processing
Human-Centered Design
Multi-Agent Systems
Robotics and Embodied AI
Future Goals:
Adaptive Multi-Agent Model: Develop an adaptive multi-agent setup to dynamically challenge ideas, enhancing the quality of user engagement and creative output.
Embodiment in Robotics: Integrate the Droodle framework into a robotic platform, allowing users to engage with an AI-driven creative experience in a tangible, interactive form.
Expanded Applications: Explore broader applications of the Droodle framework in creativity and education, including interactive storytelling and creative learning tools.