Snake-like robots are expected in disaster rescue, because they can go through narrow space that we cannot reach. Wild snakes behave various style of locomotion depending on their species or habitat. By providing feedback this feature, we must able to develop more versatile robots. We attempt to develop the robot that can locomote various terrain as the initial stage. In this study, we use model-based reinforcement learning to control snake-like robots. We demonstrate that the robot with torque information has almost same performance trained on single terrain. By adapting fast fourier fransformation (FFT) into robot’s observation, we show the difference of policy in term of frequency.