Abstract
Identifying different types of terrains is an important ability for every legged robot to achieve a stable locomotion. A variety of sensors has been applied to different robots in order to discriminate between terrains accurately. The tactile sensor has the benefits of measuring properties from terrain by physical contact between the sensor and the surface. However, the tactile sensor has rarely been utilized on quadruped robots in previous studies, and little attention has been paid to the type of sensor. There is a variety of types of tactile sensors, each with their benefits and drawbacks. The optical tactile sensor has high sensitivity, small size, light weight and low detection time, which are important properties to distinguish between different surfaces. This thesis investigates the possibility of identifying different terrains using 3D optical tactile sensors and machine learning. The measurements were retrieved from a quadruped robot developed at the University of Oslo on four different terrains. The proposed approach has the ability to classify terrains in real-time on the physical robot, and a custom segmentation method was presented for extracting desired sensor data. The segmented sensor data was the basis for creating five different feature sets and tested on five different classifiers: support vector machine, artificial neural network, naive Bayes, k-nearest neighbors, and decision tree. The experimental results demonstrated to be among the top performing approaches compared to earlier work with an accuracy of 94.8% with the support vector machine.