Generative Adversarial Networks and Data Clustering for Likable Drone Design.

Novel applications for human-drone interaction demand new design approaches, such as social drones that need to be perceived as likable by users. However, given the complexity of the likability perception process, gathering such design information from the interaction context is intricate. This work leverages deep learning-based techniques to generate novel likable drone images. We collected a drone image database (N=360) applicable for design research and assessed the drone's likability ratings in a user study (N=379). We employed two clustering methodologies: 1. likability-based, which resulted in non-likable, neutral, and likable drone clusters; and 2. feature-based (VGG, PCA), which resulted in drone clusters characterized by visual similarity; both clustered using the K-means algorithm. A characterization process identified three drone features: colorfulness, animal-like representation, and emotional expressions through facial features, which affect drone likability, going beyond prior research. We used the likable drone cluster (N=122) for generating new images using StyleGAN2-ADA and addressed the dataset size limitation using specific configurations and transfer learning. Our results were mitigated due to the dataset size; thus, we illustrate the feasibility of our approach by generating new images using the original database. Our findings demonstrate the effectiveness of Generative Adversarial Networks (GANs) exploitation for drone design, and to the best of our knowledge, this work is the first to suggest GANs for such application.