This article discusses the potential of using interactive environments in autism therapy. We specifically address issues relevant to the Aurora project, which studies the possible role of autonomous, mobile robots as therapeutic tools for children with autism. Theories of mindreading, social cognition and imitation that informed the Aurora project are discussed and their relevance to the project is outlined. Our approach is put in the broader context of socially intelligent agents and interactive environments. We summarise results from trials with a particular mobile robot. Finally, we draw some comparisons to research on interactive virtual environments in the context of autism therapy and education. We conclude by discussing future directions and open issues.