Characterizing the sense of agency in human–robot interaction based on the free energy principle

Social interaction takes various forms, ranging from cooperation to conflict. The current study focused on competition between top-down and bottom-up information processes in perception and action generation in social interaction, based on the free energy principle. In particular, it attempted to account for the sense of agency. We built a computational model for multimodal social interaction with a variational Bayes recurrent neural network, based on the free energy principle, and we evaluated the model in imitative interaction between a human and a robot. Our findings demonstrate that when prioritization of the top-down process is enhanced, the robot behaves egocentrically, leading the human more and manifesting a stronger sense of agency. Conversely, when this prioritization is reduced, the robot tends to modify its intention so that it aligns with that of its human counterpart, suggesting a weaker sense of agency. We also proposed a computational model for a multifactorial account of the sense of agency and examined the correspondence between the experimental results and the model. The study reveals the underlying mechanism of social interaction dynamics and the resultant sense of agency by conducting a rigorous analysis of the neural internal representation, as well as the behaviors of the human and the robot.