Abstract: A stable gait generation algorithm based on T-S type fuzzy learning net method is proposed in this paper. Reinforcement learning method is introduced into fuzzy network to learn the gain parameters. Few fuzzy rules are needed to formulate the nonlinear relation between the ZMP(zero moment point) curve and hip trajectory. The problem of multi-variables in continuous space is also simplified to search the independent action gains simultaneously. Simulation experiments on Luna biped robot prove its feasibility.