In traditional safety-critical systems, software diversity is of theoretical benefit but has proved to be difficult and expensive to achieve. The nature of autonomous systems and, especially, the artificial intelligence software that enables them, has the potential to change this situation. One example is that machine learning techniques, used to develop artificial intelligence software, can lower the cost of producing diverse software implementations; another is that, because it is difficult to specify precise system-level requirements for autonomous behaviour, diverse implementations can offer performance benefits. Consequently, autonomous systems may simultaneously lower the costs and increase the benefits associated with software diversity. We investigate diversity from a variety of perspectives: the concepts of diversity within a single system, or vehicle, and diversity across a fleet of vehicles are considered. We provide a simple framework that outlines different types of diversity, the kinds of benefit they provide, and the challenges associated with their use. This is intended to provide a structure with-in which principled decisions can be made about the use of software diversity in autonomous systems.