A graph neural network-based predictive control method for dynamic visual servo models

Traditional visual servo systems face three major challenges in dynamic environments: feature selection bias (poor artificial feature generalisation), model mismatch (weak dynamic scene adaptation), and real-time bottleneck (end-to-end delay > 50ms). To this end, this paper proposes a three-in-one H-MPC-GNN framework: 1) design a multi-scale edge-attention graph network (GDFE) to capture spatial and temporal associations between object motion and visual features through node-edge cooperative learning, and improve the feature selection F1-score to 0.924; 2) construct a heterogeneous model predictive control architecture with 100Hz visual layer and 10Hz control layer, and prove the asymptotic stability of the closed-loop system by time-varying Lyapunov function; 3) introduce end-to-end differentiable training, and jointly optimise the visual network parameters and the MPC weight matrix, which accelerates the speedup by 23.4% compared with the traditional two-phase training. UR5 Robot experiments show that the dynamic target tracking error of this method can be improved to 0.924 by using a multi-scale edge attention graph network (GDFE). The UR5 robot experiments show that this method has a dynamic target tracking error of <±0.1mm (83.7% lower than IBVS), an end-to-end delay of 20ms, and a partial occlusion success rate of 94.3%, which significantly improves the control accuracy and robustness in dynamic environments.