To propose a method to automatically synthesise the robot control programs by Evolutionary Computation. The problem of biped robot walking is chosen to test the proposed method. Walking motion is divided into six stages. Evolutionary Computation is carried out stage-by-stage and sub-solutions of all stages are combined to produce a final solution. Dividing the problem into searching for sub-solutions makes it easier to find the solution, because the fitness function can be set appropriately and differently in each stage. The work restricts walking locomotion to the forward walking on the flat and smooth surface with static balance. The experimental system uses simulation and the experiment with a real robot to reduce the time and to produce the solutions that work with the real robot. The result of the experiment shows that various types of gaits are achievable and the stability of biped walking is maintained continuously.