The PhD project deals with learning by observation. The objective is to learn a model of multi-modal behavior (gestures, gaze, speech etc.) for a humanoid robot given interaction traces recorded from humans performing the same task of comanipulating objects in a shared environment. The challenge is to create statistical generative models, which combine the analysis and the synthesis of human activities as well as an online adaptation of the behavior of the robot to the variations of the behavior of the human partner.