New Hybrid Fuzzy-CNN Architecture for Human-Robot Interaction

摘要

This article presents the design and the training of a new hybrid network architecture based on architectures of convolutional neural networks and a fuzzy output layer. The architecture is applied to the recognition of objects and it uses the information of distance from the point of capture of the image to the object. It seeks to address the problem of variability in the classification confidence level of a conventional convolutional network by varying the distance of the camera with respect to the object, which is presented in robot-human interaction environments, when the robot should change trajectory, so it does not collide with a person. For the tests carried out, the proposed architecture has a low value of variance and standard deviation with a value of 0.0046 and 0.068, respectively, achieving the task of gripping objects, facing interruptions of their work space, given, for example, by the interaction of a human in it.