The above sensorimotor approaches presume that an agent (robot) is equipped with an internal model of the relation between motor commands or movements and the resulting sensory states. In a behavior-based approach, the agent must learn this relation from experience. Such adaptive behavior can be achieved with neural networks, which are able to learn from examples and generalize between them. This section first gives a short overview of artificial neural networks, and then describes internal models. Finally, the last parts give an overview of existing learning paradigms for sensorimotor models: paradigms based on feed-forward networks, paradigms based on recurrent neural networks, and two extensions of the self-organizing-map algorithm (Kohonen, 1995).