Learning Deformable Linear Object Dynamics From a Single Trajectory

Abstract

The dynamic manipulation of deformable objects poses a significant challenge in robotics. While model-based approaches for controlling such objects hold significant potential, their effectiveness hinges on the availability of an accurate and computationally efficient dynamics model. This work focuses on sample-efficient learning of models to capture the dynamic behavior of deformable linear objects (DLOs). Inspired by the pseudo-rigid body method, we present a physics-informed neural ODE that approximates a DLO as a serial chain of rigid bodies interconnected by passive elastic joints. However, unlike traditional uniform spatial discretization and linear springdamper joints, our approach involves learning-based discretization and nonlinear elastic joints that characterize interaction forces via a neural network. Through real-world and simulation experiments involving DLOs with markedly different physical properties, we demonstrate the model's ability to accurately predict DLO motion.