Controllers for the locomotion of legged robots often face challenges regarding their optimization towards different objectives and different terrains. We propose an architecture that uses the information gathered in an exploration phase to adapt to a terrain with unknown characteristics. In the exploration phase virtual simulations are used to optimize the parameters of the controller in different terrains. The results of this optimization are used to identify the unknown terrain characteristics, and these values are used to select the best parameters for this particular terrain. The approach was tested in simulation, on terrains with variable friction, on an iCub robot, against a naive approach, and another where the friction was identified at random, and it clearly outper-formed in both cases.