Generative Adversarial Networks and Data Clustering for Likable Drone Design.

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Robotics

Background:

• Designing likable social drones is crucial for novel human-drone interactions.
• Perception of likability is complex and difficult to extract from interaction contexts.
• Existing drone design research lacks methods for generating likable drone aesthetics.

Purpose of the Study:

• To leverage deep learning to generate novel, likable drone images.
• To identify key visual features influencing drone likability.
• To explore the application of Generative Adversarial Networks (GANs) in drone design.

Main Methods:

• Collected a drone image database (N=360) and assessed likability ratings (N=379).
• Employed likability-based and feature-based clustering (K-means, VGG, PCA) to categorize drones.
• Utilized StyleGAN2-ADA with transfer learning to generate new drone images from a likable cluster.

Main Results:

• Identified colorfulness, animal-like features, and facial expressions as key drivers of drone likability.
• Clustering revealed distinct groups based on likability and visual features.
• Generated new drone images, demonstrating the feasibility of GANs despite dataset limitations.

Conclusions:

• Deep learning, specifically GANs, offers an effective approach for generating aesthetically pleasing and likable drone designs.
• The identified features provide actionable insights for future social drone development.
• This research pioneers the use of GANs for enhancing user perception in drone design.