Learning of soft-object models using graph neural networks

Abstract

In the field of soft-object modeling, the motion prediction of deformable linear objects (DLOs) has long been a critical challenge in robotic control and automation systems. However, DLO deformation exhibits high nonlinearity, influenced by material properties, external forces, and other factors, making accurate modeling highly challenging. Traditional modeling approaches rely on precise physical parameters, leading to high computational costs and difficulties in adapting to complex environments. As a result, data-driven methods have gradually emerged as a promising alternative for DLO modeling, among which Graph Neural Networks (GNNs) have demonstrated great potential in trajectory prediction due to their ability to model topological structures. This study proposes a GNN-based DLO trajectory prediction framework, adopting an Encode-Process-Decode structure to learn the motion dynamics of DLOs. First, a graph representation of the DLO is constructed to extract its geometric and motion features. Then, a message-passing mechanism is employed to capture both local and global dynamic characteristics and aggregate relevant information. Finally, the decoding module maps the extracted features back to the physical space to predict the future multi-step motion trajectory. Furthermore, this study explores a lightweight GNN modeling strategy. Experimental results show that by reducing the number of processing units (6–8 processors), the GNN can significantly reduce computational overhead while maintaining prediction accuracy, thereby improving inference efficiency and making it more suitable for real-time robotic control tasks.