Enhancing quadrotor trajectory prediction via hybrid-corrected TCN-MLP network

Abstract

State estimation provides the foundation for the safe and autonomous operation of quadrotors by delivering reliable pose and velocity information under noisy and incomplete sensor measurements, but it is insufficient for complex missions. As its temporal extension, trajectory prediction underpins key functions such as path planning, obstacle avoidance, and decision-making control directly. Classical estimation methods are inadequate for long-horizon prediction, while deep learning models, despite their strong nonlinear representational capacity, lack real-time correction mechanisms compatible with state estimation principles. To overcome these limitations, this paper proposes a TCN-MLP trajectory prediction model enhanced with a hybrid-correction (HC) mechanism. By dynamically incorporating partial ground-truth observations during training, the HC mechanism effectively mitigates error accumulation. Experimental results on real-world quadrotor flight datasets demonstrate that the proposed method achieves superior long-term prediction accuracy and robustness compared with mainstream baselines.