Grasping Motion Generation for Deformable Objects under Dynamic Position Changes via Variance Prediction

Because of labor shortages, robots are expected to provide work assistance in a variety of settings, including the home environment. At home we often deal with flexible objects, but flexible objects are characterized by their tendency to change position and shape. Because of this nature, data dealing with flexible objects involves uncertainty. Although deep learning has been used to perform a variety of complex tasks, the deterministic nature of conventional RNN makes it difficult to handle data with a probabilistic structure. In this study, we propose a method based on deep predictive learning that enables real-time motion generation and predicts the variance of joint angles, which facilitates learning of probabilistic structures and can handle dynamic changes. Experimental results show that the robot is able to generate motions that are adaptive to flexible objects with dynamic position changes.