Common Sensorimotor Representation for Self-initiated Imitation Learning

Internal representation is an important design decision in any imitation learning system. Actions and perceptual spaces were separate in classical AI due to the standard sense-process-act loop. Recently another representation that combines the two spaces into what we call a common sensorimotor space was inspired by the discovery of mirror neurons in animals and humans. The justification of this move is usually biological plausibility. This paper reports on a series of experiments comparing these two alternatives for self-initiated imitation tasks. The results of these experiments show that using a common sensorimotor representation allows the system to achieve higher accuracy and sensitivity. This is shown to be true (for our scenarios) even when the dimensionality of the common sensorimotor representation is higher than the dimensionality of the separate perceptual space. It also allows for an easier behavior generation mechanism and ensures reproducibility of learned behavior