Adaptive action generation against situational changes based on prediction of sensory uncertainty using neural network

Abstract

Robots working in the real environment need to respond to necessary sensory inputs. However, if the sensory inputs are not necessary for action generation, robots need to stably generate action without being affected by unnecessary sensory inputs. To realize such adaptive action generation against situational changes, robots should automatically decide how much sensory inputs are necessary for action generation. In this research, we propose a method which automatically decides the ratio of actual and predicted sensory inputs based on predicted sensory uncertainty. As a result of robot experiments, the robot with proposed mechanism could conduct adaptive action generation against situational changes.