Action learning based on developmental body perception

Abstract

The paper describes a framework for action learning in anthropomorphic robots. The key idea of the framework is that the robot voluntarily generates movements to identify its own body, and refers to the identified body in learning of fixation, reaching and grasping actions for object operation. Developmental body perception is critical to identify changing body parts in unknown or non-stationary environments. The consolidation of action learning with developmental body perception allows for continuous body awareness and anticipation of self-generated action's results. We evaluated the proposed framework in experiments with a real robot. In experiments, the robot achieved autonomous body identification, learning of fixation, reaching and gasping, and anticipation-based action planning as well as its execution in object operation tasks.