Experiments on Taking Loss Methods in Imitation Learning With Bilateral Control

Recently, motion generation technology using deep learning has been attracting attention as a method for robots to acquire motions adapted to their environment. The combination of imitation learning and bilateral control enables imitation of movements at speeds comparable to those of humans. Previous studies have investigated learning that inputs external information, such as one-hot vectors or image features, but learning to choose actions based on dynamic external information is still being studied. If a camera is used to collect external information, high-level image processing is required, resulting in heavy processing. We thought that the robot's motion generation could cover the problem without high-level image processing. Therefore, this paper examines the task of grasping an object that moves autonomously, independent of the robot, using low-resolution images. We first tried a model that takes loss based on the transformation of angle information into arm hand coordinates, whereas the previous learning model takes loss based on angle information. The results showed that, in imitation learning based on bilateral control, using hand coordinate loss caused a bias in learning the joint angle responses. It also became clear that it is important to learn the joint angle responses with equal weights.