Biped locomotion control with evolved adaptive center-crossing continuous time recurrent neural networks

We used center-crossing continuous time recurrent neural networks as central pattern generator controllers in biped robots, together with an adaptive methodology to improve the ability of the recurrent neural networks to produce rhythmic activation behaviors. The parameters of the recurrent networks are adapted or modified in run-time to reach the center-crossing condition, so the nodes get close to the most sensitive region to their input. This facilitates the evolution of the networks that act as central pattern generators to control biped structures. The robustness of the adaptive networks to produce rhythmic activation patterns was checked as well as the improvements and possibilities this adaptation may add.