The Ranger robot was designed to interact with children in order to motivate them to tidy up their room. Its mechanical configuration, together with the limited field of view of its depth camera, make the learning of obstacle avoidance behaviors a hard problem. In this article we introduce two new Particle Swarm Optimization (PSO) algorithms designed to address this noisy, high-dimensional optimization problem. Their aim is to increase the robustness of the generated robotic controllers, as compared to previous PSO algorithms. We show that we can successfully apply this set of PSO algorithms to learn 166 parameters of a robotic controller for the obstacle avoidance task. We also study the impact that an increased evaluation budget has on the robustness and average performance of the optimized controllers. Finally, we validate the control solutions learned in simulation by testing the most robust controller in three different real arenas.