Performance comparison between neural network and SVM for terrain classification of legged robot

Terrain classification of the legged robot is one of the most important objects which can determine robot's performance because surface of the fields is often extremely diverse. In the flat surface case, robot can move fast and smoothly. However, it cannot move fast in the rough terrain. Unless robot knows which terrain, robot will be falling down and slippery. Therefore, robot must know their terrain when they are walking. In this paper, we composed a 1-legged robot and terrain environment (flat, sand, and gravel) for terrain classification experiment. A load cell mounted on the 1-legged robot measures the ground reaction force and torque sensors located each of the 3-j oints measure torque. Then we present two methods for feature extraction using statistical method (Variance, Skewness, and Kurtosis) and principal component analysis (PCA) method. After that we present two methods for terrain classification such as back propagation neural network (BPNN) and support vector machine (SVM).