Perception is needed for action, not for the pure sake of the construction of abstract representations, although it does not exclude the role of internal representations for mediating complex behaviors. We think that, for the purpose of building autonomous robots, active perception requires specific recipes for three related aspects: the design of the physical sensory system, the modality and type of information extracted, and the structure and functioning of the control system. We outline a set of solutions for these three aspects and describe their implementation on a real mobile robot through a set of three different experiments using a combination of neural networks and genetic algorithms. The results show that active perception is a useful feature that is exploited by autonomous agents. The experiments show that the combination of genetic algorithms and neural networks is a feasible and fruitful technique for the development of active perception in autonomous agents.