Imitation Learning of Complex Behaviors to Humanoid Robots using Evolutionary Optimization of Neural Network of Unit Pattern Generator

In this paper, a system based on neural structures known as central pattern generator is presented which enables to acquire the required patterns to move a robot based on a demonstration training. Unit pattern generator can be divided to two subsystems, one is a rhythmic system and the other is a discrete system. The first subsystem is responsible to produce short movements and the second subsystem is responsible to produce rhythmic movements. The special learning algorithm is designed to use these unit pattern generators. Joints and limbs of robot were controlled by Kinect sensor in real time by recognition of the human body skeleton. The work steps were done in this way that the motion sequences of teacher’ body were recorded by Kinect sensor, then transmitted to the computer. These motion sequences teach some nonlinear oscillators then they reproduce motions for humanoid robot. As a result, humanoid body joints imitate the teacher movement in a real time. The main contribution of this paper is design of this learning algorithm which is able to simultaneously search for the weights and topology of the network the algorithm synchronize the neural network by coupling the neurons at the last stage.