Increasing Autonomy of Learning Sensorimotor Transformations with Dynamic Neural Fields

In this paper, we introduce a neuraldynamic architecture that enables autonomous learning of sensory-motor mappings in a closed behavioral loop. Dynamic neural fields ensure stability of perceptual and motor representations, a neural-dynamic representation of the condition-of-satisfaction autonomously terminates the current action and enables activation of the next action, triggering a transient learning process when appropriate. Sampling of the high-dimensional space of the sensory-motor mapping is facilitated by a representation of the behaviorally salient states through localized activity peaks in the visual and motor neural-dynamic representations. We present the basic concepts of our autonomous learning architecture, a robotic implementation using a dynamic vision sensor mounted on a pan-tilt unit, which enables learning in a closed behavioral loop, and demonstrate functioning of the architecture in a simple learning scenario.