Learning to walk using a recurrent neural network with time delay

Abstract

Walking based on gaits is one of the most approved methodologies for walking robot. In this paper, we develop learning strategy for walking biped robot or human based on a self made database using biomechanical capture. This system is provided by a Recurrent Neural Network (RNN) with an internal discrete time delay. The role of the proposed network is the training of human walking data by giving an estimation of the biped's next position at each time and achieve a human-like natural walking. Different architectures of RNN are proposed and tested. In Particular, a comparative study is given and the results of the RNN mixed with extended Kalman filter are illustrated. © 2013 Springer-Verlag Berlin Heidelberg.