MODELING AUTONOMOUS TRANSITIONS BETWEEN FOCUS AND MIND-WANDERING UNDER THE FREE ENERGY PRINCIPLE USING HUMANOID ROBOTS

Mind-wandering is a cognitive phenomenon that typically emerges during tasks that are either overly simple or excessively demanding. Although psychological and neuroscientific studies have investigated transitions between focused attention and mind-wandering, the computational mechanisms enabling such transitions to arise autonomously remain poorly understood. In this study, we propose a biologically grounded computational model based on the Predictive-Coding-Inspired Variational Recurrent Neural Network (PV-RNN) under the Free Energy Principle. The model introduces an online adaptive mechanism in which a meta-level prior parameter dynamically adjusts according to prediction error accumulated over a past time window. This mechanism allows the balance between top-down prediction and bottom-up sensory processing to change continuously during inference. The proposed framework was evaluated using a humanoid robot trained to predict sensory sequences generated by probabilistic transitions among multiple movement patterns. The results demonstrate that autonomous transitions between focused attention and mind-wandering emerge naturally without external intervention. Specifically, higher values of the meta-prior promote internally driven dynamics associated with mind-wandering, whereas lower values emphasize sensory-driven processing characteristic of focused attention. The observed dynamics are consistent with psychological theories of attentional fluctuation and spontaneous thought. These findings suggest a computational account of how agents may become aware of mind-wandering when accumulated prediction error exceeds a critical threshold. Overall, the proposed model provides a principled framework for investigating adaptive cognitive dynamics in artificial and biological systems.