An Adaptive Modular Recurrent Cerebellum-Inspired Controller

Animals and robots face the common challenge of interacting with an unstructured environment. While animals excel and thrive in such environments, modern robotics struggles to effectively execute simple tasks. To help improve performance in the face of frequent changes in the mapping between action and outcome (change in context) we propose the Modular-RDC controller, a bio-inspired controller based on the Recurrent Decorrelation Control (RDC) architecture. The proposed controller consists of multiple modules, each containing a forward and inverse model pair. The combined output of all inverse models is used to control the plant, with the contribution of each inverse model determined by a responsibility factor. The controller is able to correctly identify the best module for the current context, enabling a significant reduction of 70.9% in control error for a context-switching plant. It is also shown that the controller results in a degree of generalization in control.