Generating Socially Acceptable Robot Poses for Approaching Conversational Groups with Generative Adversarial Networks

Abstract

Service robots should navigate to the conversation group in a safe and socially acceptable way to interact with humans, not only to ensure safety and avoid collisions, but also to take into account human comfort. Therefore, it is crucial for robots to actively generate a target pose that conforms to social rules and satisfies human comfort feelings. At present, most methods focus on human-aware navigation and do not take into account the final pose of robots approaching the group. The target point is still manually selected by human beings, which is not conducive to improving the intelligence of robots. In this paper, we propose a target pose generation architecture based on generative adversarial networks for robots approaching conversational groups. We take into account not only the environmental information in the map, so that the robot?s target pose can avoid obstacles, but also the social etiquette of the human group, so that the robot will not disrupt the human group relationship. Our architecture uses graph convolutional neural networks to extract high-dimensional features from conversation groups, taking into account human location and gaze information. We performed a comparison experiment with a state-of-the-art pose generation model on the Cocktail Party dataset, and the results showed that our model outperformed the baseline.