Semi-Supervised Gated Recurrent Neural Networks for Robotic Terrain Classification

Legged robots are popular candidates for missions in challenging terrains due to their versatile locomotion strategies. Terrain classification is a key enabling technology autonomous legged robots, allowing them harness innate flexibility adapt the demands of operating environment. We show how highly capable machine learning techniques, namely gated recurrent neural networks, allow our target robot correctly classify terrain it traverses both supervised and semi-supervised fashions. Tests on benchmark dataset shows that time-domain classifiers well handling raw variable-length data with small amount labels outperform frequency-domain classifiers. The results own extended opens up range high-performance behaviours specific those environments. Furthermore, we unlabelled used improve significantly model.