Generating Shape Transitions of Deformable Linear Objects Using Generative Adversarial Networks

This paper presents a framework for generating shape transitions for deformable linear objects. We adopt a GAN-based approach to generate object deformations without reliance on physics simulation. We represent the deformable linear object as a point chain, and let individual shape configurations be represented as points in a low-dimensional latent space. We propose a training method for obtaining these latent representations. Then we propose a method for generating and visualizing smooth deformations. Furthermore, we propose a method for simulating how the object will deform when one of its terminal points would be moved in a given direction. We evaluate the method on the task of generating shape transitions for linear displacements of the terminal points.