Autoregressive Model Considering Low Frequency Errors in Command for Bilateral Control-Based Imitation Learning

Abstract

Recent research in the field of robotics has primarily focused on motion generation methods using imitation learning to adapt to various environments. Human-level speed motion can be imitated by combining imitation learning with bilateral control. In this approach, prediction errors accumulate in the leader because this method requires a leader state that does not exist during autonomous operation. This issue is caused by the low-frequency prediction error of the leader's responses during autonomous operation. Therefore, this paper describes a training model for predicting a leader's state by eliminating low-frequency prediction errors during autonomous operations. The proposed training model predicts the high-frequency state of the leader-follower differential. Because the low-frequency of the leader is not included in the information input and output of the proposed training model, the low frequency prediction error of the leader does not occur during autonomous operation. During our analysis, while writing letters on a paper on a slope, the previous training model failed to write letters when the paper position moved away from the robot. By contrast, only the proposed training model achieved writing at all positions in the experiment.