Multiple Soft Objects Grasping Using Bilateral Control-Based Imitation Learning

Abstract

Bilateral control-based imitation learning is an imitation learning method that predicts force commands and performs force control. Although it has advantages for contact-rich tasks, the operation frequency needs to be high to control the fine adjustment of force, and in this case, the image input is sometimes ignored. Although the authors have proposed a method of repeatedly inputting image features to each layer of a neural network, this method has been validated only for simple pick-and-place tasks and has not been tested for complex tasks. In this study, we performed a hamburger assembly task using imitation learning based on bilateral control and image feature input in each layer. By evaluating the success rate of this task, we verified the effectiveness of imitation learning based on bilateral control for tasks that require handling multiple non-rigid objects.