Dynamical Perception-Action Loop Formation with Developmental Embodiment for Hierarchical Active Inference

Kanako Esaki, Tadayuki Matsumura, Shunsuke Minusa, Yang Shao, Chihiro Yoshimura, Hiroyuki Mizuno

Abstract: To adapt an autonomous system to a newly given cognitive goal, we propose a method to dynamically combine multiple perception-action loops. Focusing on the fact that humans change their embodiment during development, the perception-action loops associated with each body part are combined. Applying the method to an end-effector movement task with a robot arm shows that the joints necessary to accomplish the target task are selectively moved in practical time. The result suggests that the robot adapts to the newly given cognitive goal and that developmental embodiment is an essential component in the design of an autonomous system.