As robots become more ubiquitous in society, they will have to learn to interact with people in socially acceptable ways. For the past six years, we have been developing techniques that enable robots to behave according to social conventions, both conversationally and spatially. The techniques involve explicit modeling of human behavior and social conventions, probabilistic reasoning about situations and the intentions of people, and explicit representation of affect and mutual interaction. We have developed several robots that embody these ideas, including GRACE, a robot that attended the National Conference on Artificial Intelligence, the roboceptionist, a joint project with the School of Drama, and a robot that dances rhythmically with children. This talk will describe our efforts in this area, focusing on the techniques that we have developed and highlighting the gap that still remains between the behavior of our robots and true social interaction.