We present a genetic programming system to evolve vision based obstacle avoidance algorithms. In order to develop autonomous behavior in a mobile robot, our purpose is to design automatically an obstacle avoidance controller adapted to the current context. Using a simulation environment, several trajectories are used to evaluate the performance of an evolving population of controllers. Two different methods are proposed and compared. In the first one (structure-free controller), we discard any a priori in the structure of the algorithm. We show that the evolved controllers behave well in the evolution environment, but have very limited generalization capabilities. In the second method (structure restricted controller), the compromise between the two antagonist functions of the robot navigation, i.e. obstacle avoidance and target reaching, is explicitely integrated in the structure of the algorithms. We show that controllers evolved with this second method are more generic. We give hints in the discussion to detect overlearning earlier in the evolution process, so as to quickly evolve controllers with good generalization abilities.