The work presented in this paper is part of the development of a robotic system able to learn context dependent visual clues to navigate in its environment. We focus on the obstacle avoidance problem as it is a necessary function for a mobile robot. As a first step, we use an off-line procedure to automatically design algorithms adapted to the visual context. This procedure is based on genetic programming and the candidate algorithms are evaluated in a simulation environment. The evolutionary process selects meaningful visual primitives in the given context and an adapted strategy to use them. The results show the emergence of several different behaviors outperforming hand-designed controllers.