In these years, many researchers are studying multi-legged robots. Since such robots have many legs, they can walk in stable; thus, they are expected to work in rough terrain. Therefore improvement of the ability for multi-legged robots to walk in rough terrain is studied. However, a walking method in rough terrain has not been developed yet. It is necessary to learn behaviors so that robots are adapted to various terrain. In this research, we aim to design a method to learn a walking policy for actual spider robot in varying rough terrain. Using the learning strategy fusion method, we acquire the walking policy where an actual spider robot learns walking task in various terrain.