FURI | Fall 2025

GNN-MPC for Path Following of Deformable Bodies

Current methodologies in robotics can handle rigid parts, but struggle with deformable structures like ropes and cables. High-fidelity models are too slow for real-time control, while hand-tuned models break when conditions change, limiting real-world application. This is fixed by training a Graph Neural Network (GNN) on Discrete Elastic Rod (DER) simulations and embedding this into a model predictive controller (MPC). The GNN captures the dynamics while the MPC guides the system to follow a prescribed path. This makes it possible to achieve real-time tracking and control with outstanding accuracy at a fraction of the computational effort, improving robustness/reproducibility.