Reinforcement learning with fuzzy evaluative feedback for a biped robot

Abstract

Proposes a fuzzy reinforcement learning algorithm for biped gait synthesis. It is based on a modified GARIC (generalized approximate reasoning for intelligent control) architecture that can accept fuzzy evaluative feedback rather than a numerical one. The proposed gait synthesizer forms the initial gait from intuitive balancing knowledge, and it is then trained by the fuzzy reinforcement learning algorithm that uses a fuzzy critical signal to evaluate the degree of success for the biped dynamic walking by means of the zero moment point. The performance and applicability of the proposed method are illustrated through biped simulation.