Learning and executing rhythmic movements through chaotic neural networks: a new method for walking humanoid robots

We propose Chaotic Neural Networks (CNN) as an alternative to other models of the Central Pattern Generation (CPG) circuits, which have been developed in the last years for robotic applications. We develop a new Matlab implementation of CNN and study their computational and functional performances. We show our results on walking humanoid robots, both in simulation and on real robots. We discuss our porting of the CNN to the on-board controller of the robot, where we verify the temporal and spatial performance. In a final comparison against CPG the CNN appear as a promising method to improve the adaptability of the robot to dynamic situations.