A Physics-Informed Dual-Branch LSTM Network for UAV Position and Attitude Estimation.

Area of Science:

• Robotics and Autonomous Systems
• Sensor Fusion and Navigation
• Deep Learning for State Estimation

Background:

• Unmanned aerial vehicles (UAVs) require accurate real-time position and attitude estimation.
• Inertial Measurement Unit (IMU) data alone suffers from error accumulation and long-term drift.
• Existing deep learning methods for inertial odometry often lack physical constraints, limiting performance.

Purpose of the Study:

• To develop a novel deep learning framework for robust IMU-only UAV state estimation.
• To mitigate error accumulation and drift in UAV navigation.
• To improve the physical interpretability and stability of deep learning-based inertial odometry.

Main Methods:

• A dual-branch physics-informed long short-term memory (DPI-LSTM) network was proposed.
• Shared temporal encoding and a dual-branch regression framework were utilized.
• Physical consistency constraints based on inertial kinematic relationships were embedded as loss functions.

Main Results:

• The DPI-LSTM network achieved a positional root mean square error (RMSE) of 0.0654 m on the UZH-FPV dataset.
• This represents a reduction of over 20% compared to state-of-the-art methods like IONet, CNN-LSTM, and RoNIN.
• Ablation studies confirmed the benefits of physical constraints and the dual-branch architecture for improved stability.

Conclusions:

• The proposed kinematically constrained DPI-LSTM framework offers a viable solution for IMU-only position and attitude estimation.
• The model provides continuous and physically consistent navigation data essential for UAV digital twins and precision agriculture.
• This research advances the field of inertial navigation for autonomous systems.