Architectures for functional imagination

Imagination can be defined broadly as the manipulation of information that is not directly available to an agent's sensors. However, the topic of imagination raises representational, physiological, and phenomenological issues that cannot be tackled easily without using the body as a reference point. Within this framework, we define functional imagination as the mechanism that allows an embodied agent to simulate its own actions and their sensory consequences internally, and to extract behavioural benefits from doing so. In this paper, we present five necessary and sufficient requirements for the implementation of functional imagination, as well as a minimal architecture that meets all these criteria. We also present a taxonomy for categorising possible architectures according to their main attributes. Finally, we describe experiments with some simple architectures designed using these principles and implemented on simulated and real robots, including an extremely complex anthropomimetic humanoid.