Object Recognition Using Multiple Neural Networks and Force Sensing

Abstract

Object recognition is an important research area in robotics. This paper proposes two methods, voting method and softmax method that collect reaction data by pressing objects and combines five simple neural networks to recognize these objects. In the experiment, each data sample was divided into up to five sections and each section was used to train or test its corresponding model, respectively. Five optimal models are formed by selecting proper hyperparameters. In the voting method, five models recognized objects respectively; the label appeared the most is considered as the final object label. In the softmax method, five softmax function output vectors of these five models are averaged as the final vector; the final object label is calculated based on this final vector. In this research, the robot pressed 49 household objects, about 110 times per object. The results show that, the robot can improve its recognition accuracy by combining five neural network models and the softmax method are move effective.