Neuro-based controller for push recovery behavior under external perturbations in biped robot

This paper presents neuro-based push recovery controller applied in humanoid biped robot in order to keep the stability with minimum energy required. There are three motion patterns in human behavior when it gets external perturbation, those are ankle behavior, hip behavior, and step behavior. We propose the new model of modular recurrent neural network (MRNN) for performing online learning system in each motion behavior. MRNN consists of several recurrent neural networks (RNNs) working alternately depending on the condition. MRNN performs online learning process of each motion behavior controller independently. The aim of push recovery controller is to manage the motion behavior controller by minimizing the energy required for responding to the external perturbation. This controller selects the appropriate motion behavior and adjusts the gain that represent the influence of the motion behavior to certain push disturbance based on behavior graphs which is generated by adaptive regression spline. We applied the proposed controller to the humanoid robot that has small footprint in open dynamic engines (ODE). Experimental result shows the effectiveness of the push controller stabilizing the external perturbation with minimum energy required.