Generalization of a mental rotation skill in humanoid robots

Abstract

This work demonstrates how generalization ability can be integrated into a neural network model of mental rotation. The model was validated with a physical humanoid robot, the iCub, as simulated participants. The results confirm that the proposed model is capable of solving a mental rotation task consisting of a number of unseen stimuli. Furthermore, characteristic of response time profiles and error rates replicates the same fashion as found in human participants. Mechanisms underlie the successes are forward model training and matching processes, both are independent of objects' identity. This work could benefit robotic applications e.g., planning, decision-making, in which the results of any actions can be seen before really performing them.