I. Introduction
In interaction situations involving groups of people that are standing and involved in conversation, a vital ability is to join the group by approaching it. When robots approach these free-standing conversational groups [1], they should adopt socially acceptable paths in order not to make individuals in the group feel uncomfortable, for example, due to violating their personal boundaries (see Fig. 1). Due to the importance of these approach behaviors for robots that have social roles, including mobile companion robots and delivery robots in social environments [2], [3], [4], recently a number of navigation methods and experiments have been devised in relation to robot approach behaviors [5]. Although these methods and experiments involve safe and socially acceptable paths, collectively they have shortcomings that limit their utility as a full solution for robot navigation into groups: a) most research focuses on robot behaviors that approach an individual human, b) most of the cases involve groups of humans that are totally static within the conversational group i.e. they do not change body orientations and positions at all, while the positions and orientations of the individuals in free-standing groups are not totally static over time, c) the experimental studies typically do not propose computational models, d) existing computational models are only evaluated in simulation with handcrafted features and specific measurements. To overcome the aforementioned difficulties, we present a data-driven method to generate robot approach behaviors into a conversational group. To the best of our knowledge, no other work has previously used data-driven methods for this purpose. The main contribution of the paper is our introduction of a Long-Short Term Memory network (LSTM) based Generative Adversarial Network (GAN) with a novel group interaction module that fuses body position and orientation information of individuals in a group to generate socially acceptable paths for a mobile robot when it approaches a ‘quasi-dynamic’ conversational group. Our model is trained and evaluated on a semi-synthetic dataset of safe and socially acceptable paths. We show that it outperforms baseline methods via the GAN and our novel group interaction module, leading to efficient and safe trajectories for robots in free-standing conversational groups.
A sample from our dataset (refer to Fig. 5). Members of the conversational group (red circle) are quasi-dynamic, changing body positions and orientations. The robot initially plans a path (red curve) to approach and join the group without interrupting the conversation while approaching from the front (left). However, the two people in the group change positions as well as body orientations (right) and the robot thus changes its path accordingly (green curve) to continue to approach the group from the front.
‘Quasi-dynamic’ refers to slight changes in the body positions and orientations of people in a free-standing conversational group.