Surface classification enhancement using estimated power consumption data

Choosing the right control strategy has always been an important aspect of autonomous localization and navigation. One of the ways of doing it is to identify the surface characteristics and assign the appropriate controller. The most prominent ways of recognizing the surface type are either to do it visually or using proprioceptive signals induced by robot–environment interactions. However, these interactions are very complex and to obtain the full perspective the classifier should make the decision based on data from several diverse sources. A potential source, which has not been extensively studied in the literature, is power consumption measured at the locomotion devices. The question which this study aims to answer is whether this signal carries enough unique information to boost the classification result in a significant way. The analysis incorporates usage of two different machine learning algorithms, namely a convolutional neural network and gradient boosted trees, to determine the difference in classification results between training on vibration data only and on a combination of vibration and estimated power consumption data. The results show that the gap between these training datasets is meaningful and the additional information about power consumption might positively impact the classification outcome.