Parallax-based Imitation Learning for Insertion Tasks with Uncertainties of Hole Position and Wire Pose

Abstract

Standard insertion machines require pre-determined component position and posture. If they change every time, we must solve this problem. Most conventional methods attempted to solve this task by identifying the position and posture. However, these methods require a multi-step strategy following the handmade rule. This paper proposes an imitation learning method to automate the wire insertion task with uncertainties in position and posture. The proposed model learns the motion policy through human demonstrations and maps image data to the robot's action in a single step. Moreover, the model considers the parallax of the stereo images for accurate insertion. In addition, the model outputs the insertion action and recovery action to recover from insertion failures. However, the standard data collection method cannot collect recovery actions, and manual labeling of action classes is essential. This paper proposes a novel data collection method called "Labeling with Human Intervention (LHI)" to tackle this problem. This method automatically generates action labels and collects recovery action with human intervention. We conducted real-space insertion tests and found that our approach achieved 97.2% (35/36).