Determining Distance to an Object and Type of its Material Based on Data of Capacitive Sensor Signal and Machine Learning Techniques

Abstract

Capacitive proximity sensors allow detecting the presence of nearby objects without any physical contact with them, as well as in poor visibility conditions in the presence of dust or smoke. Tactile sensors in robotic devices provide additional information about the environment to the robot control system, and particularly serve as a feedback element for an object gripping system or a robot's coordination and gait system. In this work, tactile sensors were used to determine the distance from the object to the sensor and to recognize the type of object material. An array of capacitive sensors of four cells was used. We tested the effectiveness of seven machine learning methods (support vector machine (SVM), decision tree, naive Bayesian classifier (NBC), random forest, k-nearest neighbors method, gradient boosting, neural network model (Keras, ReLU activation function)) in the problem of determining the distance to an object, as well as its type of material. According to the results of the experiment, the most accurate method for recognizing the type of object material was the k-nearest neighbors method (96.9%), and the most accurate method for determining the distance to an object based on the output signal of capacitive sensors and the object material was a random forest (78.7%).