The task of physically assisting humans requires from robots the ability to adapt in many different ways: to changes in space of the human movement, to changes in the speed of the human, to changes in the environment, etc. This paper presents recent research on teaching robots how to interact with humans and to adapt to different circumstances.The approach presented here is based on Imitation Learning and Probabilistic Movement Representations. In particular, this paper explains the concept of a Mixture of Interaction Primitives to learn interactions from multiple unlabeled demonstrations and to deal with nonlinear correlations between the interacting partners. Furthermore, a method to compute reactions to human movements executed at different speeds is presented. A number of experiments with a lightweight robotic arm illustrate applications of the presented methods.