Generating Social-Aware Locations for a Robot in a Human Group by Image Generation Using Adversarial Learning

摘要

With the advance of deep learning techniques, social robots can have more powerful perception and interaction capabilities. However, the problem of finding a socially aware standing location for the robot to join a conversation group is not well addressed. Thus, we propose a generative-based and image-based approach to generate a social-aware group formation to obtain the possible locations for the robot. Furthermore, to overcome the problem of formulating human comforts, we try to leverage human behaviors with the concerns of human comforts when joining the conversation group. We utilize a self-supervised technique to generate this kind of human experience from the real-world dataset. Through extensive experiments, we show that the proposed method outperforms the social force method by 62% with respect to data from human experiences. In addition, our approach also provides controllable parameters to generate the location with the required features using the GAN noise vector.