Socially-aware navigation planner using models of human-human interaction

Abstract

In this paper, we revisit a real-time socially-aware navigation planner which helps a mobile robot to navigate alongside humans in a socially acceptable manner. This navigation planner is a modification of nav core package of Robot Operating System (ROS), based upon earlier work and further modified to use only egocentric sensors. The planner can be utilized to provide safe as well as socially appropriate robot navigation. Primitive features including interpersonal distance between the robot and an interaction partner and features of the environment (such as hallways detected in real-time) are used to reason about the current state of an interaction. Gaussian Mixture Models (GMM) are trained over these features from human-human interaction demonstrations of various interaction scenarios. This model is both used to discriminate different human actions related to their navigation behavior and to help in the trajectory selection process to provide a social-appropriateness score for a potential trajectory. This paper presents an evaluation done in simulation while utilizing data from real human interactions.