Abstract: Objective: In traditional Extended Kalman Filter (EKF)-based GNSS/SINS integration, IMU bias is typically modeled as a slowly varying first-order Gauss-Markov process. However, rotor UAVs operating in complex environments often face high-dynamic maneuvers and GNSS signal occlusion. Under these conditions, low-cost MEMS IMU biases exhibit significant instability and random jumps, rendering standard assumptions invalid and leading to inaccurate bias estimation and rapid divergence of positioning errors. To address these limitations, this study proposes a multi-task perception learning method that fuses UAV flight state information to achieve online bias compensation.Methods: The proposed method utilizes a Multi-Layer Perceptron (MLP) with a hard parameter sharing strategy to construct a multi-task learning bias prediction network (MTL-BP). This architecture simultaneously performs motion state recognition and IMU bias prediction. To resolve the ambiguity between hovering and constant-velocity flight states, time-frequency joint features are constructed by fusing IMU time-domain integration with the Power Spectral Density (PSD) of the Y-axis accelerometer. Subsequently, the six motion state probabilities output by the classification task are concatenated with IMU features into a 13-dimensional vector for the bias prediction task. The resulting reliable bias values are used for error feedback correction in a tightly coupled GNSS/SINS system.Results: Real-world flight experiments were conducted in a mountainous environment using a DJI M600 UAV. Results indicate that the introduction of frequency-domain features effectively resolved the confusion between hovering and constant-velocity flight, improving motion state recognition accuracy from 91.07% to 93.64%. Regarding bias prediction, the MTL-BP model demonstrated dynamic tracking capabilities superior to single-task models, significantly suppressing prediction jumps during maneuvers. In terms of positioning performance, the 3D Root Mean Square Error (RMSE) was reduced by 24.4% during intense maneuvering. Furthermore, during a simulated 20-second GNSS outage, the method effectively curbed error divergence, reducing the 3D positioning error from 1.44 m to 0.85 m, an improvement of 41.0%.Conclusion: The proposed multi-task learning approach reveals the intrinsic correlation between UAV motion states and IMU biases. It achieves precise system error compensation in complex dynamic environments and significantly enhances the robustness of GNSS/SINS integrated navigation systems against aggressive maneuvers and signal interruptions, demonstrating substantial potential for high-precision UAV navigation.