In evolutionary robotics, we evaluate individuals by placing them in an initial configuration in the environment, and then measure their fitness over a period of time. The choice of initial configuration has a direct impact on the fitness of an individual and thereby also the overarching evolutionary process. In this paper, we propose the concept of dynamic initial configurations, which is an initial configuration that is neither random nor fixed, but develops dynamically in response to the evolutionary process. As an example we have implemented a competitive co-evolutionary algorithm where initial configurations and controllers are evolved together to solve an obstacle avoidance task of a mobile robot. We show that, while a evolutionary approach taken from literature consistently fails, the co-evolutionary approach succeeds in 22 out of 25 runs. This example demonstrates the benefit of dynamic initial configurations, but more work is needed to establish if the concept generalizes to more complex tasks, environments and morphologies.