MULTIMODAL AFFECT MODELING AND RECOGNITION FOR EMPATHIC ROBOT COMPANIONS

摘要

Affect recognition for socially perceptive robots relies on representative data. While many of the existing affective corpora and databases contain posed and decontextualized affective expressions, affect resources for designing an affect recognition system in naturalistic human-robot interaction (HRI) must include context-rich expressions that emerge in the same scenario of the final application. In this paper, we propose a context-based approach to the collection and modeling of representative data for building an affect-sensitive robotic game companion. To illustrate our approach we present the key features of the Inter-ACT (INTEr-acting with Robots-Affect Context Task) corpus, an affective and contextually rich multi-modal video corpus containing affective expressions of children playing chess with an iCat robot. We show how this corpus can be successfully used to train a context-sensitive affect recognition system (a valence detector) for a robotic game companion. Finally, we demonstrate how the integration of the affect recognition system in a modular platform for adaptive HRI makes the interaction with the robot more engaging.