MGpi: A Computational Model of Multiagent Group Perception and\n Interaction

Abstract

Toward enabling next-generation robots capable of socially intelligent\ninteraction with humans, we present a $\\mathbf{computational\\; model}$ of\ninteractions in a social environment of multiple agents and multiple groups.\nThe Multiagent Group Perception and Interaction (MGpi) network is a deep neural\nnetwork that predicts the appropriate social action to execute in a group\nconversation (e.g., speak, listen, respond, leave), taking into account\nneighbors' observable features (e.g., location of people, gaze orientation,\ndistraction, etc.). A central component of MGpi is the Kinesic-Proxemic-Message\n(KPM) gate, that performs social signal gating to extract important information\nfrom a group conversation. In particular, KPM gate filters incoming social cues\nfrom nearby agents by observing their body gestures (kinesics) and spatial\nbehavior (proxemics). The MGpi network and its KPM gate are learned via\nimitation learning, using demonstrations from our designed $\\mathbf{social\\;\ninteraction\\; simulator}$. Further, we demonstrate the efficacy of the KPM gate\nas a social attention mechanism, achieving state-of-the-art performance on the\ntask of $\\mathbf{group\\; identification}$ without using explicit group\nannotations, layout assumptions, or manually chosen parameters.\n