Data Augmentation for Continual Learning of Fast and Smooth Imitation Motions Using Model Predictive Control

Research aimed at enabling robots to acquire not only repetitive and simple tasks but also complex and delicate tasks is receiving significant attention to expand the use of robots to support and replace workers. Imitation learning is one of the promising approaches to enable robots to learn complex human skills with minimal learning cost. However, since the motion data generated by human-operated robots serves as the reference motion for the robot, it is difficult for the robot to acquire fast and smooth motions that surpass human operation. In this paper, we propose a data augmentation method to generate faster and smoother motion data by adjusting the robot’s motion output from the trained model. By utilizing model predictive control (MPC) for motion adjustment, it becomes possible to balance trackability to the original motion and smoothness of the motion through the design of the evaluation functions and constraints. The robot can perform tasks efficiently and robustly by continuously learning augmented motion data that has been optimized using MPC. We demonstrate through experiments on object picking and placing task that higher-quality motion data generated in the real-world.