Natural Data-driven Approaching Behaviors of Humanoid Mobile Robots for F-Formations

In this work we address the problem of approaching a group of persons based on a data-driven model, which is gathered in real-world human interaction. Our model estimates an adequate pose for approaching groups of different sizes, providing a mobile robot with the ability to decide how to approach a group. The model is capable of automatically detecting and distinguishing individuals from groups of humans, although only approaches to groups will be focused. From group detection, it computes the 2D poses of the group's elements and, based on the poses, estimates an approaching pose to join it. The model uses an F-Formation oriented representation of the world that allows estimating an approaching pose to groups regardless of the environment and the arrangement of the elements of the group in space. The model was tested in simulations and real-world experiments. Results show that the model was fully capable of estimating, in both tested scenarios, human-like and socially acceptable approaches to groups.