Robot imitation learning based on cyber-graphic model

Cybernetic graphic model (CGM) based on visual observation for robot imitation learning was proposed, which regards the behaviors as stage control procedure, and uses sequence of behavioural primitive to express and execute. Firstly, theoretical precondition analysis proved that differential motion of terminal executor could be used as the behavioral primitives. Then architecture of CGM was described. Eventually, learning method of CGM was proposed, which consisted of three steps: (1)accumulating and instantaneous correlation function was used to make segmental criterions of sequences to build CGM architecture. (2)behavioral sequence was represented to CGM nodes in trajectory level based on gradient descent under arc length constraint. (3)Radial basis function (RBF) network was used to functionalize the node parameters to boost the generalization ability of model. The experiments of two behavioral instances, which cover different levels and types of behaviors, on two types of robots prove that CGM is extremely advanced and powerful in generalization of imitation learning.