Creating controllable, responsive avatars is an important problem in computer games and virtual environments. Recently, large collections of motion capture data have been exploited for increased realism in avatar animation and control. Large motion sets have the advantage of accommodating a broad variety of natural human motion. However, when a motion set is large, the time required to identify an appropriate sequence of motions is the bottleneck for achieving interactive avatar control. In this paper, we present a novel method for training avatar behaviors from unlabelled motion data in order to animate and control avatars at minimal runtime cost. Based on machine learning technique, called Q-teaming, our training method allows the avatar to learn how to act in any given situation through trial-and-error interactions with a dynamic environment. We demonstrate the effectiveness of our approach through examples that include avatars interacting with each other and with the user.