Embodied Categorization of Spatial Environments on the Basis of Proprioceptive Data

This article describes research on the usability of proprioceptive data for spatial categorization. A first experiment was carried out using the 8-legged robot SCORPION, which provides us with a multitude of proprioceptive sensors. We present how these signals can be used in our biomimetic approach to learn on which substrates the robot is moving. Results are presented from runs carried out on our experimental indoor test bed, which contains different substrates, e.g., sand, rocks and gravel. To classify the proprioceptive sensor data, a supervised learning approach based on the growing cell structure learning algorithm is used. The gained results are discussed and next steps are motivated.