Controlling the Sense of Agency in Dyadic Robot Interaction: An Active Inference Approach

This study investigated how social interaction among robotic agents changes dynamically depending on individual sense of agency. In a set of simulation studies, we examine dyadic imitative interactions of robots using a variational recurrent neural network model. The model is based on the free energy principle such that interacting robots find themselves in a loop, attempting to predict and infer each other's actions using active inference. We examined how regulating the complexity term to minimize free energy during training determines the dynamic characteristics of networks and affects dyadic imitative interactions. Our simulation results show that through softer regulation of the complexity term, a robot with stronger agency develops and dominates its counterpart developed with weaker agency through tighter regulation. When two robots are trained with equally soft regulation, both generate individual intended behavior patterns, ignoring each other. We argue that primary intersubjectivity does develop in dyadic robotic interactions.