Machine learning based locomotion strategy selection for a legged wheeled hybrid quadruped

Abstract

In this work we present the Nazarbayev University Hybrid Quadruped (Fig. 1) – four legged and four wheeled mobile robot platform capable of navigating in both configurations avoiding the shortcomings of robots with single locomotion mode. The major novelty of the robot is the implementation of the supervisory controller which selects a locomotion mode associated with particular terrain types based on the classification of depth images acquired by a RGB-Depth camera. The accuracy of the terrain classification for the testing database in a five class terrain recognition problem (Level Ground, Nontraversable, Stair Down, Stair Up, Uneven terrain types) is 96.7% percent. Majority voting filter technique is used to increase the reliability of the terrain recognizer decisions.