Research on the control strategy of robot imitation learning based on KL divergence

Abstract

Imitation learning is a significant means of human learning, but also the main research field in the mechanism of bionic robot. This paper focuses on the imitation learning strategies of robot in the framework of probabilistic model. Discrete teaching data are used as training samples of Gaussian process to acquire demonstration trajectory. RBF neural network is adopted to express imitation control strategy. The imitation trajectory with imitation control strategy which contains unknown parameters is modeled by Gaussian process. KL divergence is constructed with the probability distribution of demonstration and imitation trajectory, and gradient descent method is used to minimize the KL divergence so as to seek the optimal strategy of imitation. Then the imitation task is learned gradually by updating the optimal strategy to imitative robot. The swing behavior of the articulated robot arm is used as the simulation task of imitation learning, and the result of simulation experiments demonstrates the effectiveness of the robotic control strategy for imitation learning based on KL divergence and RBF neural network.