Detail publikace

Abstrakt

Proprioception and tactile sensing in insect-like legged robots is a fast, illumination insensitive and biologically inspired way of ground perception. In this thesis, 14 virtually generated terrains are classified based on the mentioned sensor types for a simulated version of hexapod robot AMOS II. A feedforward neural network framework equipped with a novel network pruning algorithm has been developed for classification. We observe over 92% classification accuracy on deterministic terrain data and 72% on manually noised data. The pruning algorithm removes unimportant synapses (generally more than 90%) from a fully-connected network, while the classification accuracy does not drop significantly. The number of input neurons is reduced by 65%, resulting in the minimal network structure for the classification problem. A theory of using minimal structures for feature selection is proposed. The thesis outcome consists of a minimal neural network capable of terrain classification based on selected features of proprioceptive and tactile sensory signals.