EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Robust State Estimation for Maneuvering UAVs: A Set Transformer-Based Adaptive UKF Approach

Junjie Zhou, Xi Zhao, Junhao Liu, Xinyu Xu, Aoxuan Yang, Fan Pu

IEEE Access · 2026 · 14 · 97118-97142

DOI: 10.1109/access.2026.3703071 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Junjie Zhou und 5 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Reliable state estimation is critical for Uncrewed Aerial Vehicles (UAVs) navigating complex urban environments. However, traditional filters and recent recurrent neural network (RNN) based estimators often fail or exhibit severe lag during aggressive maneuvers due to rigid process noise covariance (Q) assumptions and sequence-induced hysteresis. To address these limitations, this paper proposes the A-KIT UKF, a novel framework bridging deep learning with classical Bayesian estimation. By utilizing a Set Transformer to process sliding windows of measurement residuals as permutation-invariant, unordered sets, the architecture instantaneously identifies maneuver signatures. This fundamentally eliminates the temporal lag inherent in traditional adaptive algorithms and LSTM-based networks, enabling real-time, axis-decoupled covariance scaling. Practically, this ensures that UAVs can execute aggressive collision-avoidance maneuvers in obstacle-rich urban canyons without sacrificing vertical stability or suffering from trajectory overshoot. High-fidelity simulations involving stochastic wind and high-g maneuvers confirm the method’s absolute superiority. By treating sliding windows of residuals as permutation-invariant sets, this architectural shift eliminates the temporal lag inherent in LSTM-based networks. Evaluated under identical experimental constraints, the A-KIT UKF significantly reduces the position Root Mean Square Error (RMSE) to 0.8010 m during sustained high-g coordinated turns (an 88.2% improvement over the standard EKF). Furthermore, the network’s anisotropic scaling explicitly prevents coupled multi-axis oscillations, improving attitude estimation accuracy by up to 55.2% in transient maneuvers. With a highly lightweight parameter footprint and an ultra-low neural network inference latency of 0.60 ms (with a total filter step time of 2.61 ms) on an Intel Core i7-13700H CPU @ 2.40GHz and an NVIDIA GeForce RTX 4060 GPU, this approach provides a robust, real-time, and efficient solution for next-generation Uncrewed Aerial Vehicle (UAV) vehicles operating in GPS-degraded, obstacle-rich environments.

Zitieren

APA

Zhou, J., Zhao, X., Liu, J., Xu, X., Yang, A., & Pu, F. (2026). Robust State Estimation for Maneuvering UAVs: A Set Transformer-Based Adaptive UKF Approach. IEEE Access, 14, 97118-97142. https://doi.org/10.1109/access.2026.3703071

Vancouver

Zhou J, Zhao X, Liu J, Xu X, Yang A, Pu F. Robust State Estimation for Maneuvering UAVs: A Set Transformer-Based Adaptive UKF Approach. IEEE Access. 2026;14:97118-97142. doi:10.1109/access.2026.3703071

BibTeX
@article{Zhou2026,
  author = {Zhou, Junjie and Zhao, Xi and Liu, Junhao and Xu, Xinyu and Yang, Aoxuan and Pu, Fan},
  title = {Robust State Estimation for Maneuvering UAVs: A Set Transformer-Based Adaptive UKF Approach},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {97118--97142},
  doi = {10.1109/access.2026.3703071}
}
Kostenlos starten →