LSTM-GCN Hybrid Architecture for Model Predictive Control of Deformable Linear Objects

Manipulation tasks involving flexible objects, including deformable linear objects (DLOs), are widely and critically applied in manufacturing, medical, and service industries. With the continuous progress in embodied intelli-gence, the precise control of the DLO shape using dual robotic arms has attracted considerable discussion. However, controlling DLOs is challenging due to their high degrees of freedom, strong nonlinearity, and complex dynamic characteristics. In addressing the dynamic manipulation and control problem of deformable linear objects, this study simultaneously considers both the spatial and temporal characteristics of DLOs. A model predictive approach is proposed by integrating LSTM and GCN algorithms, which learns from simulation data to predict DLO deformation in real time. In practical operation, a ridge regression-based real-time feedback correction mechanism is used to compensate for model prediction errors, and a model predictive controller is established to optimize the robotic motion online. Finally, simulation experiments are performed, and comparisons with existing methods demonstrate this approach's advantages in control precision and response speed, with terminal error reduced by 3.2% and settling time decreased by 9.5%.