One common problem of dynamic robot simulations is the accuracy of the actuators’ behavior and their interaction with the environment. Especially when simulating legged robots which have optimized gaits resulting from machine learning, manually finding a proper configuration within the high-dimensional parameter space of the simulation environment becomes a demanding task. In this paper, we describe a multi-staged approach for automatically optimizing a large set of different simulation parameters. The optimization is carried out offline through an evolutionary algorithm which uses the difference between the recorded data of a real robot and the behavior of the simulation as fitness function. A model of an AIBO robot performing a variety of different walking gaits serves as an example of the approach.