Avoiding the world model trap: An acting robot does not need to be so smart!

We propose several examples of robotics applications where interactions with the environment play an important role, and where no modelization of the robot or of the world is needed. We argue that, in general, this helps to simplify the complexity of the control system, and illustrate this claim through several examples. We propose different neural networks which allow clustering of scattered objects by several robots, learning of obstacle avoidance and learning of target retrieval.