Exploring surface detection for a quadruped robot in households

Abstract

Surface recognition is essential for legged robots because they need to maintain their dynamic balance on a regular or uneven terrain. The accelerometer is a widely-used tool for this purpose, but the quadruped Sony AIBO does not have such a high-end sensor compared to the latest state-of-art developments. Past works focused on attaching replacement sensors to the robot dog as well as collecting many samples for machine learning methods although some studies did not address this issue at all. This paper focuses on improvements with sensor fusion of built-in sensors to recognize wider variety of surfaces and get similar or better accuracy than earlier experiments. The combined features are based on the accelerometer and paw sensors to make the recognition more robust and a naive Bayes classifier achieves 85-91% accuracy for different locomotion speeds. Evaluation suggests that this method can reduce the data collection time for training samples dramatically and it is suitable for practical applications.