To mitigate error accumulation and long-term drift in unmanned aerial vehicle (UAV) position and attitude estimation using purely inertial measurement unit (IMU) data, this paper presents a dual-branch physics-informed long short-term memory (DPI-LSTM) network incorporating shared temporal encoding, a dual-branch structured regression framework, and physical consistency constraints. The model employs a long short-term memory (LSTM)-based temporal encoder to extract temporal features from IMU time-window sequences. Established inertial kinematic relationships are embedded into the dual-branch LSTM framework as loss constraints, providing physics-based regularisation to guide the network during training. By modelling translational and rotational states separately through the position and attitude branches, the model improves stability and physical interpretability while retaining the advantages of task decoupling. Systematic experiments were conducted on the University of Zurich First-Person View (UZH-FPV) Drone Racing dataset, and comparisons were made with traditional inertial navigation methods and representative deep learning-based inertial odometry approaches. The experimental results indicate that the proposed model demonstrates a measurable reduction in positional root mean square error (RMSE) on the evaluated test sequences, decreasing the RMSE to 0.0654 m, which represents a reduction of more than 20% when compared with inertial odometry network (IONet), convolutional neural network–long short-term memory (CNN–LSTM), and robust neural inertial navigation (RoNIN). Further ablation studies and cross-sequence evaluation indicate that the physical consistency constraints and the dual-branch architecture contribute to improved position estimation stability under the evaluated benchmark sequences. The proposed kinematically constrained framework provides a viable IMU-only position and attitude estimation module, laying the groundwork for future UAV digital twin and precision-agriculture applications where continuous and physically consistent position and attitude information is required.
Liang et al. (Mon,) studied this question.