Motion Generation Based on Contact State Estimation Using Two-stage Clustering

Abstract

Many contact-rich tasks in factories are still executed manually because it is often difficult for robots to adapt to variable environmental conditions. To incorporate the adaptability of humans to the environment into robots, a machine learning technique called learning from demonstration (LfD) has been well studied. A popular approach for automating complex tasks involves selecting the appropriate action from several segmented motions, called movement primitives. In this study, we propose a method for autonomously selecting a recovery action and correcting the trajectory when the task is determined to have failed based on force. Although conventional technologies can divide motions into time series, they are unable to recognize the presence of failures in detail in response to slight environmental variations. Therefore, we propose a two-stage clustering method, which consists of time segmentation of trajectories and labeling of segmented motions, to recognize failures and generate of recovery actions in response to the failures. The proposed method was able to perform the task even in cases of a 20 mm position error by accurately selecting recovery actions and correcting the trajectory.