Trajectory Correction for Glue-Application Task by a Robot Arm Using Force and BiLSTM

In this paper, we propose a trajectory correction method that utilizes force and bidirectional long short-term memory (BiLSTM) to reproduce the delicate glue-application task performed by a human on a robot. In the system we have developed, the operator initially performs the glue-application task using a teaching device. The trajectory and force are recorded during the task. If the robot follows the recorded trajectory, it can be expected to perform the same glue-application task as the human. However, large errors in the force occur owing to the deflection of the robot and individual differences in the application tool. To minimize this problem, our previous study used force and iterative learning to correct the trajectory. Although the error was reduced, running the robot multiple times for iterative learning took a long time. The more types of teaching are learned, the longer time is required. To solve this problem, BiLSTM was used in this study. To make BiLSTM learn correction equivalent to iterative learning, the trajectory and force recorded by the teaching device and robot during the glue-application task were used. Achieving this should allow deep learning to correct the robot’s trajectory, even for an unknown teaching move, because of its better generalization performance compared with other machine learning algorithms. In this study, we investigated whether this is possible.