NARLE: Neurocognitive architecture for the autonomous task recognition, learning, and execution

Robots controlled by the state of the art cognitive architectures are still far behind animals in their capabilities to learn complex skills and autonomously adapt to unexpected circumstances. The neurocognitive architecture proposed in this paper addresses the problem of learning and execution of hierarchical behaviors and complex skills. Learning is addressed both on the level of individual elementary behaviors and goal-directed sequences of actions. The proposed architecture comprises a Dynamic Neural Fields (DNFs) implementation of the low-level elementary behaviors and a Functional System Network (FSN) tying these behaviors in goal-directed sequences. The DNF framework enables a continuous, dynamical representation of perceptual features and motor parameters, which may be directly coupled to the robot’s sensors and motors. Attractor states and instabilities of the DNFs account for segregation of cognitive states and mark behaviorally relevant events in the continuous flow of sensorimotor dynamics. The FSN, in its turn, comprises dynamical elements that can be arranged in a multilayered network by a learning process, in which new layers and elementary behaviors are added on demand. In our architecture, the FSN controls adaptation processes in the already acquired neural-dynamic elementary behaviors, as well as formation of new elementary behaviors. Combination of the DNF and FSN frameworks in a neurocognitive architecture NARLE enables pervasive learning both on the level of individual behaviors and goal-directed sequence, contributing to the progress towards more adaptive intelligent robotic systems, capable to learn new tasks and extend their behavioral repertoire in stochastic real-world environments.