MagGAT: Structure-Aware Spatiotemporal Graph Attention for Motor-Induced Magnetic Interference Compensation in Quadruped Robots

Permanent-magnet synchronous motors enable precise motion control in compact robots, but their leakage fields can corrupt onboard magnetometer measurements. Physics-based compensation is effective under fixed geometric conditions, whereas quadruped locomotion continuously changes the motor–sensor relationship. We propose MagGAT, a structure-aware spatiotemporal graph attention network that estimates compensated magnetic field signals from motor states and near-field measurements. A boom-mounted far-field magnetometer provides a pseudo-reference for supervised training and field-level evaluation. MagGAT combines a global–local dual-branch graph attention module for multi-motor spatial coupling with a gated recurrent unit for temporal disturbance modeling. Experiments on a Unitree Go2 quadruped used synchronized near-field and far-field measurements, joint angles, angular velocities, and commanded torques collected in indoor and outdoor environments. Across three runs with different random seeds, MagGAT achieved mean residual three-axis magnetic field RMSE values of 87.72–89.41 nT across the three test sequences and outperformed a physics-based fixed-parameter model, a static neural network method, and a temporal convolutional baseline. Ablation studies confirmed that both spatial graph modeling and temporal modeling contributed to compensation accuracy. These results support structure-aware spatiotemporal learning for suppressing motor-induced magnetometer interference on compact legged robots.