This paper presents an algorithm of a robust controller used for the stabilization of biped robot walking gait. The robot is described by a kajita's model with norm bounded uncertainties to ensure a more realistic numerical model. The proposed robust dynamic controller relies on the use of predictive control theory (MPC) and the resolution of a convex optimization problem with Linear Matrix Inequalities (LMI) at every sampling period. The generated control law allows a realtime working robot even in rough ground or unknown environment.