부산대학교 도서관

Demonstration learning based on Probabilistic Movement Primitives (ProMP) has been widely used in robotics skill learning. For trajectory planning in traditional ProMP, the sequential online learning method is adopted. In other words, only one data point is considered at each time, and the model parameters are updated correspondingly. This usually leads to the problem that as the number of new data points to be fitted increases, old points that could be fitted accurately by the model are now not fitted accurately. In this paper, we demonstrate that the degree of uncertainty in the prediction distribution gradually decreases as the number of observed data points increases, which is responsible for the occurrence of the above phenomenon. To solve this problem, we propose a weight combination algorithm. Every point to be fitted is processed one by one and the basis functions that fall within the highly correlated range with the point to be fitted are involved in the regression operation. Finally, the weight vector components corresponding to these basis functions are concatenated and combined to obtain the complete weight vector. We mathematically prove that the new algorithm is better than the traditional online algorithm. At the end of this paper, the simulation experiments are given to prove the rationality of the new algorithm and the accuracy higher than the traditional ProMP.