Sound-Based Terrain Classification for Multi-modal Wheel-Leg Robots

Abstract

The mobile robot will move continually with the ground in the various unstructured environments, and it will inevitably be affected by the terrain geometry and type (physical property). As a result, terrain detection and classification skills are critical and demand extra attention in order to assure the reliability of robot control and navigation. For terrain analysis of specific tasks, previous work has always used vision-based non-contact sensors or proprioceptive contact-based sensors (e.g., IMU and force sensors), even though vision is not robust against environmental changes and signals acquired by IMU or force sensors are not rich enough. In this paper, we proposed a contact-based terrain classification method with a novel acoustic sensing modality that can provide much richer contact information for robot-terrain interactions and mobile mobility. To this end, we use a multi-functional test-bed with 6 different terrain types for data collection. We then use 4 machine learning algorithms to assess and handle numerous sensory signals (audio and force). The experiment results show that the acoustic signal can reach a great classification accuracy of more than 98%, which is much higher than the force signal. The comparison of multiple terrain types and robot locomotion modality indicates the robustness and effectiveness of our proposed sound-based terrain classification method.