Dynamic Recurrent Neural Network for Biped Robot Equilibrium Control: Preliminary Results

The purpose of the research addressed in this paper is to develop a real time neural control algorithm for the balance of a biped robot. Our approach is based on dynamic recurrent neural networks and dynamic backpropagation through time algorithm. The neural architecture and its learning process are validated on the control of the ROBIAN biped torso. The neural controller described is trained to compensate, by the torso’s joint motions, applied external perturbations. The algorithm is embedded in the real time electronic unit of the robot and online learning is achieved. The learning behavior and the control performances are the preliminary results presented in this paper. These experimental results show the ability and efficiency of the proposed approach.