Terrain Classification and Negotiation with a Walking Robot

This paper describes a walking robot controller for negotiation of terrains with different traction characteristics. The feedback is based on three perception systems. The purpose of the presented research is to enhance the autonomy of the walking robot. The information about the class of the terrain allows the robot to work in the real world scenarios more efficiently. In the presented work twelve types of the ground were tested. Suitability of each type of the perception system for characterizing the terrain class was checked. Namely, vision, depth and tactile sensors were used. In each case as a classifier the Support Vector Machines were utilized. The separate classification results from each sensor were combined to obtain better precision and recall in the ground classification process. The information about the terrain type was fed into robot controller to adapt the robot gait parameters. The goal was to achieve good balance between the speed of the movement and the vibration caused by the bounciness and the irregularities of the terrain. The paper begins with the description of the experimental setup. Next, the classification results for each sensor used are presented. Then, the rules of combining classifiers were tested. Finally, the robot gait controller was proposed and evaluated.