Sequential Motion Primitives Recognition of Robotic Arm Task via Human Demonstration Using Hierarchical BiLSTM Classifier

Abstract

Learning from demonstration (LfD) is an intuitive teaching technology without extensive programming for an operator. In recent LfD research, machine vision is usually used to capture the human-robot interaction. However, it's not reliable during the machining process. In this letter, a novel intuitive high-level kinesthetic teaching technology is proposed by reconstructing the motion information recorded from human-guided demonstrations. A hierarchical BiLSTM-based machine learning algorithm is proposed in this letter to recognize and segment motion primitives according to the therblig definition. A hybrid sensing interface is used to record and extract the motion features, consisting of the velocity profile, force/torque, and gripper information. The motion features are then used to classify into the target motion primitive by the proposed classifier. The experimental results and comparisons with the state-of-the-art algorithm show that the proposed method can correctly and efficiently synthesize the recorded motion features into a motion primitive sequence. Finally, the recognition results of real-world tasks show that the proposed algorithm can be used to reconstruct the human-guided task and further used to command a KUKA robot. The experimental results of the reconstructed trajectory show that a real-world task can represent and maintain the accuracy in 2.37 mm using the proposed algorithm.