Acoustics Based Terrain Classification For Legged Robots

Legged robots offer a more versatile solution to traversing outdoor uneven terrain compared to their wheeled and tracked counterparts. They also provide a unique opportunity to perceive the terrain-robot interactions by listening to the sounds generated during locomotion. Legged robots such as hexapod robots produce rich acoustic information for each gait cycle which includes the foot fall sounds and feet pushing on the terrain (support phase), as well as the sounds produced when the feet travel through the air (stride phase). Interpreting this information to perceive the terrain it is traversing makes available another valuable sensing modality which can feed in to higher level systems to facilitate robust and efficient navigation through unknown terrain. We present an online real-time terrain classification system for legged robots that utilise features from the acoustic signals produced during locomotion. A 32-dimensional feature vector extracted from acoustic data recorded using an on-board microphone was fed in to a multi-class Support Vector Machine (SVM). The SVM was trained on 7 different terrain types and the results of the experimental evaluations are presented. The system was implemented using the Robotic Operating System (ROS) for real-time terrain classification. A classification time-resolution of 1 s was achieved by capturing acoustic signals of two steps, and the results show a true positive rate ( sensitivity) of up to 92.9%. We also present a noise subtraction technique which removes servo noise and improves the sensitivity up to 95.1%.