Adaptive joint trajectory generator based on a chaotic recurrent neural network

The aim of this paper is to introduce a scalable and adaptable joint trajectory generator based on a recurrent neural network. As main application we target highly redundant kinematic structures like humanoid and multi-legged robotic systems. The network architecture consists of a set of leak integrators which outputs are limited by sigmoidal activation functions. The neural circuit exhibits very rich dynamics and is capable to generate complex periodic signals without the direct excitation of external inputs. Spontaneous internal activity is possible thanks to the presence of recurrent connections and a source of Gaussian noise that is overlapped with the signals. By modulating the internal chaotic level of the network it is possible to make the system exploring high-dimensional spaces and therefore to learn very complex time sequences. A preliminary set of simulations demonstrated how a relatively small network composed of hundred units is capable to generate different motor paths which can be triggered by exteroceptive sensory signals.