We propose a method to automatically assess group openness by observing interactions within socially occupied spaces. Openness refers to a group’s predisposition to allow external parties to join a social encounter. Robots need the capability of assessing openness to ensure social comfort and improve first impressions, saving time and resources by avoiding unwelcomed interactions. We extract social signals from gestures, posture, space, and constellations using skeleton data from depth sensors. We collected 88 unique around 5 minutes group interactions involving 82 Japanese participants across two scenarios: one with an external focus of attention (a poster illustration on the wall, Poster scenario) and another without one (standing conversation, No Poster scenario). We trained a linear classifier optimised with a Stochastic Gradient Descent on raw data, as well as aggregated bodily motion and spatial cues with a full and a half temporal overlap by implementing a sliding window temporal analysis across observation windows, demonstrating its capability to assess group openness accurately. Our approach was evaluated across temporal analysis, scenarios comparison, and group sizes, reaching 79% peak prediction accuracy and an F1 score of 0.8201 for the No Poster scenario and 78% with 0.8147 F1 score for the Poster scenario, outperforming a human baseline of 50% and 70% respectively. Feature importance scores suggest groups display different cues for openness in each scenario. Important cues for the No Poster scenario included tightness, group size, body crunching, right arm extension and mouth covering. For the Poster scenario, relevant cues were arm extension, distance to the wall, gap between members and tightness. These findings contribute to understanding group openness prior to engagement, improving robots’ social intelligence in successfully approaching human groups.
Sosa-León et al. (Tue,) studied this question.