4족 보행 로봇의 험지 보행을 위한 지형인식 알고리즘 개발

Abstract

Terrain classification algorithms were proposed for the purpose of properly changing the gait patterns according to the ground conditions. The quadruped robot is needed to verify the proposed algorithms, but it is very complex and costly. The one-leg platform was thus developed to simplify the experiment and to construct the four indoor terrains. When the one-leg platform walks on the four terrains, the ground reaction force (GRF) is acquired from the load cell mounted on the end effector. The GRF data were used as the input data of the terrain classification algorithms. Features were extracted to reduce the noise and to improve the calculation cost. The moment analysis (MA) and principal component analysis (PCA) methods were used as feature extraction methods as they are very powerful tools and are being used for various applications in the recognition field. The extracted feature data and class data were used as the input data of the supervised learning algorithms. Backpropagation neural network (BPNN) and support vector machine (SVM) are the most popular learning algorithms. The success rate of these algorithms depends on the combination of the feature extraction method and the supervised learning algorithm. The feature dimensions extracted through the MA and PCA methods are also critical factors in the performance of the algorithms. Thus, terrain classification was performed using four algorithms (MA-BPNN, PCA-BPNN, MA-SVM, and PCA-SVM), to compare their performances. Finally, the best combination and the optimal feature dimension were determined. The success rate of the best case was over 90%. Comparatively speaking, it demonstrated better performance in terms of terrain classification compared to the wheel-type mobile robot.