In this paper, we demonstrate how an artificial neural network (ANN) based controller can be synthesized for a complex task through hierarchical evolution and composition of behaviors. We demonstrate the approach in a task in which an e-puck robot has to find and rescue a teammate. The robot starts in a room with obstacles and the teammate is located in a double T-maze connected to the room. We divide the rescue task into different sub-tasks: (i) exit the room and enter the double T-maze, (ii) solve the maze to find the teammate, and (iii) guide the teammate safely to the initial room. We evolve controllers for each sub-task, and we combine the resulting controllers in a bottom-up fashion through additional evolutionary runs. We conduct evolution offline, in simulation, and we evaluate the highest performing controller on real robotic hardware. The controller achieves a task completion rate of more than 90% both in simulation and on real robotic hardware.