Shared Perception in Human-Robot Interaction

Interaction can be seen as a composition of perspectives: the integration of perceptions, intentions, and actions on the environment two or more agents share. For an interaction to be effective, each agent must be prone to “sharedness”: being situated in a common environment, able to read what others express about their perspective, and ready to adjust one’s own perspective accordingly. In this sense, effective interaction is supported by perceiving the environment jointly with others, a capability that in this research is called Shared Perception. Nonetheless, perception is a complex process that brings the observer receiving sensory inputs from the external world and interpreting them based on its own, previous experiences, predictions, and intentions. In addition, social interaction itself contributes to shaping what is perceived: others’ attention, perspective, actions, and internal states may also be incorporated into perception. Thus, Shared perception reflects the observer's ability to integrate these three sources of information: the environment, the self, and other agents. If Shared Perception is essential among humans, it is equally crucial for interaction with robots, which need social and cognitive abilities to interact with humans naturally and successfully. This research deals with Shared Perception within the context of Social Human-Robot Interaction (HRI) and involves an interdisciplinary approach. The two general axes of the thesis are the investigation of human perception while interacting with robots and the modeling of robot’s perception while interacting with humans. Such two directions are outlined through three specific Research Objectives, whose achievements represent the contribution of this work. i) The formulation of a theoretical framework of Shared Perception in HRI valid for interpreting and developing different socio-perceptual mechanisms and abilities. ii) The investigation of Shared Perception in humans focusing on the perceptual mechanism of Context Dependency, and therefore exploring how social interaction affects the use of previous experience in human spatial perception. iii) The implementation of a deep-learning model for Addressee Estimation to foster robots’ socio-perceptual skills through the awareness of others’ behavior, as suggested in the Shared Perception framework. To achieve the first Research Objective, several human socio-perceptual mechanisms are presented and interpreted in a unified account. This exposition parallels mechanisms elicited by interaction with humans and humanoid robots and aims to build a framework valid to investigate human perception in the context of HRI. Based on the thought of D. Davidson and conceived as the integration of information coming from the environment, the self, and other agents, the idea of "triangulation" expresses the critical dynamics of Shared Perception. Also, it is proposed as the functional structure to support the implementation of socio-perceptual skills in robots. This general framework serves as a reference to fulfill the other two Research Objectives, which explore specific aspects of Shared Perception. For what concerns the second Research Objective, the human perceptual mechanism of Context Dependency is investigated, for the first time, within social interaction. Human perception is based on unconscious inference, where sensory inputs integrate with prior information. This phenomenon helps in facing the uncertainty of the external world with predictions built upon previous experience. To investigate the effect of social interaction on such a mechanism, the iCub robot has been used as an experimental tool to create an interactive scenario with a controlled setting. A user study based on psychophysical methods, Bayesian modeling, and a neural network analysis of human results demonstrated that social interaction influenced Context Dependency so that when interacting with a social agent, humans rely less on their internal models and more on external stimuli. Such results are framed in Shared Perception and contribute to revealing the integration dynamics of the three sources of Shared Perception. The others’ presence and social behavior (other agents) affect the balance between sensory inputs (environment) and personal history (self) in favor of the information shared with others, that is, the environment. The third Research Objective consists of tackling the Addressee Estimation problem, i.e., understanding to whom a speaker is talking, to improve the iCub social behavior in multi-party interactions. Addressee Estimation can be considered a Shared Perception ability because it is achieved by using sensory information from the environment, internal representations of the agents’ position, and, more importantly, the understanding of others’ behavior. An architecture for Addressee Estimation is thus designed considering the integration process of Shared Perception (environment, self, other agents) and partially implemented with respect to the third element: the awareness of others’ behavior. To achieve this, a hybrid deep-learning (CNN+LSTM) model is developed to estimate the speaker-robot relative placement of the addressee based on the non-verbal behavior of the speaker. Addressee Estimation abilities based on Shared Perception dynamics are aimed at improving multi-party HRI. Making robots aware of other agents’ behavior towards the environment is the first crucial step for incorporating such information into the robot’s perception and modeling Shared Perception.