Generalizing Multimodal Policies Learned in a Human-Robot Interaction