Terrain Classification for Bipedal Robots: A Comparative Study

This paper presents a comparison of machine learning classification techniques used to identify nine different terrains based on data acquired from force, current, position, and inertial sensors of the NAO humanoid robot. We will be comparing four different classification techniques based on Support Vector Machines, K Nearest Neighbor, Naive Bayes, and Random Forest. Additionally, we will be exploring Manual Feature Reduction, Principle Component Analysis, and Linear Discriminant Analysis to identify which feature reduction method performs better in this application. This paper aims to compare the performance of different classification techniques as well as their performance vis-a-vis different subsets of the available sensors. This will allow us to choose the most suitable technique for our case given the large number of sensors and type of data we collected. At the end, we present the machine learning technique with the best feature reduction method achieving the highest accuracy.