Using concepts and tools of embodied cognitive science, we investigate the implications of embedding neural networks in a physical structure, the body of a robot. Embedding a neural network in a body provides constraints that can be exploited for learning. We show that the constraints are given by the environment and object properties, the agent's morphology, the agent's motor system and specific ways of interacting with the objects. We argue that designing embedded neural networks implies (a) understanding these constraints, and (b) exploiting them, i.e., designing neural networks such that they-one way or other-incorporate the constraints. This in turn results in cheap and simple networks that are suited for the task environment, and have real-time responses. Moreover, this constraint-based approach provides new perspectives on two fundamental problems of cognitive science: focus-of-attention and object constancy. The main arguments are illustrated with a series of case studies with simulated and physical mobile robots that are controlled by hand-designed as well as evolved neural networks.