On the role of feature and signal selection for terrain learning in planetary exploration robots

Increasing the terrain awareness of planetary exploration rovers is one key technology for future space robotics to successfully accomplish long-distance and long-duration missions. In contrast to most of the existing algorithms that use visual or depth data for terrain classification, the approach presented in this study tackles the problem using proprioceptive sensing, for example, vibration or force measurements. The underlying assumption is that these signals, being directly modulated by the terrain properties, are well descriptive of a given surface. Therefore, terrain signature can be inferred via learning algorithms that are trained on either the signals directly or a signal-derived feature set. Following the latter approach, first, a physics-based signal augmentation process is presented that aims at maximizing the information content. Then, a feature selection algorithm based on a scoring system and an iterative search is developed to decrease the computational cost while preserving high classification accuracy. The resulting most informative feature subspace can be used to train a support vector machine (SVM) classifier. For comparison, the time histories of the selected proprioceptive signals are used to train a deep convolutional neural network (CNN). Results obtained from real experiments using the SherpaTT rover confirm that proprioceptive sensing is effective in predicting terrain type with an accuracy higher than 90% for both algorithms in generalization tasks. When the two learning approaches are contrasted in extrapolation problems, for example, predicting observations acquired at previously unseen velocity or terrain, CNN outperforms the standard SVM. Furthermore, CNN holds the additional advantage of learning features automatically from signal spectrograms, reducing the need of a priori knowledge at the expense of higher computational efforts.