A Novel Teacher-Assistance-Based Method to Detect and Handle Bad Training Demonstrations in Learning From Demonstration

Abstract

Learning from demonstration (LfD) assists robots to derive a policy from a training set to execute a task. The training set consists of training demonstrations collected from a human who executes the same task. However, due to the human’s varied skill level, the quality of the training set may be bad, which will affect the accuracy of the derived policy. To solve this problem, this article proposes a novel method to improve the quality of the training set. This method includes two steps, namely detecting and handling bad training demonstrations in the training set. In the detecting step, a reference set containing reference demonstrations is provided by a human teacher. Based on the reference set, we calculate the influence of each training demonstration on the policy derivation. If the influence is negative, the corresponding training demonstration is bad. Afterward, in the handling step, we calculate the proportion of the negative influence with respect to the overall influence and reduce the proportion by iteratively removing bad training demonstrations until it is less than a threshold. The results show that the accuracy of a policy derived from the improved training set increases with up to 19.30%, which verifies the effectiveness of our method.