GRAB-HAI : génération des comportements réciproquement adaptatifs pour les interactions humain-agent
Jieyeon Woo · theses.fr (ABES) · 2023
Information is transferred from one person to another via communication. Through this transfer, we convey our thought and intentions via multimodal signals such as words, gestures, and prosody. This exchange of signals is a two-way process of sending and receiving where the behaviors of the interlocutors adapt to each other. Such adaptation is continuous, dynamical, and reciprocal which we refer to as reciprocal adaptation. Adapting to others allows interactions to be engaging and effective. Endowing such capacity to embodied agents, Socially Interactive Agents (SIAs) and robots, can make them more social and engaging and perceived as natural and human-like. Nevertheless, this endowment is a challenging task. The agents need to know how to adapt as both a speaker or a listener while emitting behaviors related to its own speech synchronized over its modalities, intrapersonal relationship, and with its interlocutor’s behaviors, interpersonal relationship. The central focus of this thesis is to develop an adaptive SIA with reciprocal adaptation capabilities. We propose computational models, ASAP and HI2-ADAM, to render SIA’s adaptive behaviors as both speaker or a listener. ASAP generates adaptive and continuous behavior using multimodal signal information from its user and itself by modeling the interpersonal relationship between them. HI2-ADAM captures the reciprocal adaptation and intrapersonal relationship in an explicit way by modeling each modality history of each interlocutor and learning from the relation between these different histories. As it is important for agents to act as interaction partners and continuously adapt their behaviors in real time, we create a real-time interactive and adaptive agent, IAVA system, and provide new measures, reciprocal adaptation measures, for the evaluation of human-agent interaction quality.