Physics-aware Learnable State Space Model for UAV Trajectory Prediction

Unmanned aerial vehicle (UAV) trajectory prediction plays an important role in autonomous flight control, path planning, and obstacle avoidance decision-making. However, the accumulation of errors in multi-step autoregressive predictions, lack of physical consistency, and real-time constraints make existing pure data-driven models limited for engineering deployment. Therefore, this paper proposes a physics-aware learnable state space prediction framework (PLSSP). Unlike directly performing black-box regression on positions, this paper models UAV states as a joint physical state vector consisting of position, velocity, and acceleration. By using residual modeling, only the dynamic perturbation terms are learned, and a maneuver-aware selective state update mechanism is incorporated to enable adaptive modeling of the changes in flight dynamics. Additionally, the propagation behavior of multi-step prediction errors is analyzed for stability, and the model's prediction performance and engineering metrics are systematically evaluated using real PX4 flight log data. Experimental results show that, while ensuring real-time inference efficiency, the proposed method outperforms several mainstream time series prediction models in terms of Average Displacement Error (ADE) and long-term prediction stability.