LAUSR

Inertial Measurement Units (IMUs), comprising gyroscopes and accelerometers, are fundamental for motion estimation in navigation and robotics. However, their performance is often degraded by nonlinear and time‐varying errors, such as bias drift, scale‐factor deviations, and sensor noise. Traditional compensation methods based on linear assumptions or static models struggle to address the dynamic and correlated nature of these errors, limiting real‐time calibration robustness. To address this challenge, we propose a transformer‐based framework for gyroscope error compensation, which dynamically models temporal dependencies and nonlinear error characteristics using self‐attention mechanisms. Our approach incorporates a sliding window to exploit historical sensor data and applies a geometrically constrained loss function defined on the SO(3) Lie group, ensuring physically consistent orientation estimates. Comprehensive experiments on the public EuRoC and Technical University of Munich data sets demonstrate that our method achieves a mean orientation error of 1.56°, outperforming state‐of‐the‐art approaches including robust 3‐D orientation estimation with a single particular IMU (9.46°), DenoiseIMU (2.09°), and Gyro‐Net (1.49°). Additionally, our framework reduces the Absolute Trajectory Error (ATE) by 45.6% (average 0.070 m) and the Relative Pose Error by 48.2% (average 0.0043 m) compared with established baselines. These results highlight the effectiveness and robustness of our method, particularly in challenging scenarios with rapid motion and low‐texture environments. Overall, our transformer–based approach significantly enhances the reliability and accuracy of IMU‐based systems, offering a promising solution for autonomous navigation and related applications.