This thesis moves closer to the vision of social machines in the real world, making fundamental contributions along the three dimensions of sensing, interpreting and anticipating nonverbal behaviour in social interactions. First, it advances the state of the art in human visual behaviour sensing, proposing a novel unsupervised method for eye contact detection in group interactions by exploiting the connection between gaze and speaking turns. Furthermore, the thesis makes use of mobile device engagement to address the problem of calibration drift that occurs in daily-life usage of mobile eye trackers. Second, this thesis improves the interpretation of social signals by proposing datasets and methods for emotion recognition and low rapport detection in less constrained settings than previously studied. In addition, it for the first time investigates a cross-dataset evaluation setting for emergent leadership detection. Third, this thesis pioneers methods for the anticipation of eye contact in dyadic conversations, as well as in the context of mobile device interactions during daily life, thereby paving the way for interfaces that are able to proactively intervene and support interacting humans.