Hidden partitioning of a visual feedback-based neuro-controller

Robotic controllers take advantage from neural network learning capabilities as long as the dimensionality of the problem is kept moderate. This paper explores the possibilities offered by the combination of several neural networks to design more complex modular controllers. This modularity is based on an internal partitioning of the problem. The partitioning must remain hidden, and should not affect the controller's interface or functioning, including during its adaptation phases. We introduce a bi-directional architecture to derive the learning rules of the modules. The neurocontroller is trained globally, based on the interactions of the system with its environment, as one would do for a single network. The approach is illustrated on a robotic reaching application. Several partitioning variants of the neuro-controller are discussed and compared.