高精度定位為自主移動機器人執行任務規劃、導航、避障之重要基礎。於全球導航衛星系統（Global Navigation Satellite System, GNSS）訊號受限之環境中，超寬頻（Ultra-Wideband, UWB）可提供高精度測距資訊，而慣性量測單元（Inertial Measurement Unit, IMU）則能提供短時間內連續之運動資訊。透過 UWB 與 IMU 融合，可互補 UWB 量測易受遮蔽與非視距傳播影響，以及 IMU 長時間積分易累積漂移之限制。然而，UWB 測距誤差易受視距（Line-of-Sight, LOS）與非視距（Non-Line-of-Sight, NLOS）傳播條件影響，並在遮蔽與多路徑環境下呈現非平穩特性。因此，不同基站於不同時間與場域條件下之量測可信度往往有顯著不同，若未能即時反映此差異，將影響定位結果之準確度與穩健性。
基於上述考量，本研究提出一套結合緊耦合 UWB/IMU 誤差狀態卡爾曼濾波器（Error-State Kalman Filter, ESKF）之長短期記憶網路（Long Short-Term Memory, LSTM）自適應量測協方差估測方法。所提方法以 UWB 量測之訊號品質指標與 IMU 之量測資訊建立組合時間序列，作為 LSTM 之輸入，以學習時間序列與測距誤差之關係，進而估測各 UWB 基站之量測雜訊標準差，並轉換為緊耦合 ESKF 量測更新所需之各基站量測雜訊協方差，使量測更新過程能依據各基站之即時可靠度調整其測距量測權重。本研究以實地蒐集所得之真實實驗資料建構模型所需之訓練集，並結合同步參考軌跡標註各基站之測距誤差，作為模型之監督式學習標籤。完成訓練後，模型被整合至緊耦合 ESKF 架構中，以 IMU 作為預測來源，並直接引入各基站 UWB 測距作為量測更新資訊。相較於先行解算 UWB 位置後再融合之鬆耦合架構，緊耦合架構可保留各基站之原始資訊，更適合應用於基站可視數量不足之場域。濾波器狀態向量包含位置、速度、姿態、IMU 偏置與各基站 UWB 測距動態偏差，並結合 z 軸位置、z 軸速度與姿態平面約束之偽量測更新，以抑制垂直方向漂移與姿態誤差對定位結果之影響。
為驗證所提方法之可行性與成效，本研究於大樹遮蔽訊號場景與金屬物體多路徑場景中進行定位實驗，並部署於自主割草機平台上。由實驗結果可觀察到，所提方法可降低定位誤差，同時提高系統於遮蔽與多路徑環境下之穩健性；不確定度分析亦顯示，所估測之量測雜訊協方差可反映實際測距誤差分布，使緊耦合濾波器能依據各基站測距資訊之可靠度合理調整更新權重，並在可視基站數量少於四顆時仍能穩定運作。本研究完成由資料蒐集、模型訓練、即時協方差估測、緊耦合濾波器整合至實際平台驗證之完整流程，證明所提方法可提升 GNSS 受限環境中 UWB/IMU 融合定位之精度與可靠性。

High-precision localization serves as an important foundation for autonomous mobile robots to perform task planning, navigation, and obstacle avoidance. In environments where Global Navigation Satellite System (GNSS) signals are subject to errors, Ultra-Wideband (UWB) can provide high-precision ranging information, while an Inertial Measurement Unit (IMU) can provide continuous motion information. By fusing UWB and IMU, it is possible to compensate for the susceptibility of UWB measurements to blockage and non-line-of-sight propagation, as well as the tendency of IMU to accumulate drift due to long-term integration. However, UWB ranging errors are easily affected by Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) propagation conditions, and exhibit non-stationary characteristics under blockage and multipath environments. As a result, the measurement reliability of different anchors often varies significantly across different times and field conditions; if this variation is not reflected in real time, the accuracy and robustness of the fused localization results will be affected.
Based on the above considerations, this study proposes an adaptive measurement-noise covariance estimation method that combines a tightly coupled UWB/IMU Error-State Kalman Filter (ESKF) with a Long Short-Term Memory (LSTM) network. The proposed method constructs a combined time series from UWB signal-quality indicators and IMU measurement information, which is used as the input to the LSTM to learn the relationship between the time series and ranging errors, thereby estimating the measurement-noise standard deviation of each UWB anchor. This is then converted into the measurement-noise covariance of each anchor required for the tightly coupled ESKF update, allowing the ranging weight of each anchor to be adjusted during the measurement update process according to its real-time reliability. This study constructs the training dataset using real experimental data collected in the field, and labels the ranging error of each anchor based on a synchronized reference trajectory, serving as supervised learning labels for the model. After training, the model is integrated into the tightly coupled ESKF framework, using IMU as the prediction source and directly incorporating UWB ranging from each anchor as measurement update information. Compared with a loosely coupled architecture that first solves for the UWB position before fusion, the tightly coupled architecture preserves the raw information from each anchor, making it more suitable for scenarios where the number of visible anchors is insufficient. The filter state vector includes position, velocity, attitude, IMU bias, and the dynamic ranging bias of each UWB anchor, and incorporates pseudo-measurement updates based on z-axis position, z-axis velocity, and attitude-plane constraints to suppress the effects of vertical drift and attitude error on the localization results.
To verify the feasibility and effectiveness of the proposed method, this study conducts localization experiments under a tree-blocked scenario and a metallic-object multipath scenario, and deploys the system on an autonomous lawn mower platform. The experimental results show that the proposed method reduces localization error while improving system robustness under blockage and multipath conditions; uncertainty analysis further shows that the estimated measurement-noise covariance reflects the actual distribution of ranging errors, enabling the tightly coupled filter to reasonably adjust update weights according to the reliability of ranging information from each anchor, and to operate stably even when fewer than four anchors are available. This study completes the whole process from data collection, model training, and real-time covariance estimation, to the integration of the tightly coupled filter and validation on an actual platform, demonstrating that the proposed method can improve the accuracy and reliability of UWB/IMU fused localization in GNSS-limited environments.

[1] Dalian Haoru Technology Co., Ltd., RTLS1-LD600(-I) User Manual, Dalian Haoru Technology Co., Ltd., Jul. 2022, version 1.0, Document No. LD600(-I)_UserManual.

[2] P. Nguyen-Thanh, M.-Y. Cho, C.-L. Chang, and M.-J. Chen, "Short-term three-phase load prediction with advanced metering infrastructure data in smart solar microgrid based convolution neural network bidirectional gated recurrent unit," IEEE Access, vol. 10, pp. 68 686–68 699, 01 2022.

[3] Y. Yu, X. Si, C. Hu, and J. Zhang, "A review of recurrent neural networks: LSTM cells and network architectures," Neural Computation, vol. 31, no. 7, pp. 1235–1270, 2019.

[4] NovAtel Inc., OEM7720 Product Sheet, Hexagon | NovAtel, July 2023, d21633 Version 12.

[5] u-blox AG, ZED-F9P Series Product Summary, u-blox, 2024, uBX-17005151 - R17.

[6] G. M. Mendoza-Silva, J. Torres-Sospedra, and J. Huerta, "A meta-review of indoor positioning systems," Sensors, vol. 19, no. 20, p. 4507, 2019.

[7] W. C. S. S. Simões, G. S. Machado, A. M. A. Sales, M. M. de Lucena, N. Jazdi, and V. F. de Lucena, "A review of technologies and techniques for indoor navigation systems for the visually impaired," Sensors, vol. 20, no. 14, p. 3935, 2020.

[8] S. G. Leitch, Q. Z. Ahmed, W. B. Abbas, M. Hafeez, P. I. Lazaridis, P. Sureephong, and T. Alade, "On indoor localization using wifi, BLE, UWB, and IMU technologies," Sensors, vol. 23, no. 20, p. 8598, 2023.

[9] D. Feng, C. Wang, C. He, Y. Zhuang, and X.-G. Xia, "Kalman-filter-based integration of IMU and UWB for high-accuracy indoor positioning and navigation," IEEE Internet of Things Journal, vol. 7, no. 4, pp. 3133–3146, 2020.

[10] W. You, F. Li, L. Liao, and M. Huang, "Data fusion of UWB and IMU based on unscented Kalman filter for indoor localization of quadrotor UAV," IEEE Access, vol. 8, pp. 64 971–64 981, 2020.

[11] M. F. R. Al-Okby, S. Junginger, T. Roddelkopf, and K. Thurow, "UWB-based real-time indoor positioning systems: A comprehensive review," Applied Sciences, vol. 14, no. 23, p. 11005, 2024.

[12] F. Wang, H. Tang, and J. Chen, "Survey on NLOS identification and error mitigation for UWB indoor positioning," Electronics, vol. 12, no. 7, p. 1678, 2023.

[13] I. Guvenc, C.-C. Chong, and F. Watanabe, "NLOS identification and mitigation for UWB localization systems," in 2007 IEEE Wireless Communications and Networking Conference, 2007, pp. 1571–1576.

[14] J. Dong, Z. Lian, J. Xu, and Z. Yue, "UWB localization based on improved robust adaptive cubature Kalman filter," Sensors, vol. 23, no. 5, p. 2669, 2023.

[15] M. Flissi, K. Rouabah, and S. Atia, "Cumulative adaptive extended Kalman filter for robust UWB positioning in indoor environments," International Journal of Communication Systems, vol. 38, no. 10, p. e70127, 2025.

[16] H. Yang, Y. Wang, S. Xu, J. Bi, H. Jia, and C. Seow, "Ultra-wideband ranging error mitigation with novel channel impulse response feature parameters and two-step non-line-of-sight identification," Sensors, vol. 24, no. 5, p. 1703, 2024.

[17] X. Wen, J. Yang, J. Tian, and T. Chao, "UWB-inertial fusion localization algorithm based on error-state Kalman filter in GNSS-denied environments," in Advances in Guidance, Navigation and Control - Proceedings of 2024 International Conference on Guidance, Navigation and Control Volume 18, ser. Lecture Notes in Electrical Engineering, L. Yan, H. Duan, and Y. Deng, Eds. Springer Science and Business Media Deutschland GmbH, 2025, pp. 239–248.

[18] W. Zhu, R. Zhao, H. Zhang, J. Lu, Z. Zhang, B. Wei, and Y. Fan, "Improved indoor positioning model based on UWB/IMU tight combination with double-loop cumulative error estimation," Applied Sciences, vol. 13, no. 18, p. 10046, 2023.

[19] L. Zhang, D. Sidoti, A. Bienkowski, K. R. Pattipati, Y. Bar-Shalom, and D. L. Kleinman, "On the identification of noise covariances and adaptive Kalman filtering: A new look at a 50 year-old problem," IEEE Access, vol. 8, pp. 59 362–59 388, 2020.

[20] A. Chhabra, J. R. Venepally, and D. Kim, "Measurement noise covariance-adapting Kalman filters for varying sensor noise situations," Sensors, vol. 21, no. 24, p. 8304, 2021.

[21] N. Xu, M. Guan, and C. Wen, "A survey on ultra wide band based localization for mobile autonomous machines," Journal of Automation and Intelligence, vol. 4, no. 2, pp. 82–97, 2025.

[22] J. Cano, Y. Ding, G. Pages, E. Chaumette, and J. Le Ny, "A robust kalman filter based approach for indoor robot positionning with multi-path contaminated UWB data," in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5.

[23] P. T. Karfakis, M. S. Couceiro, D. Portugal, and R. Cortesão, "UWB aided mobile robot localization with neural networks and the EKF," in 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2022, pp. 93–99.

[24] U. Albertin, M. Martini, A. Navone, and M. Chiaberge, "Adaptive robot localization with ultra-wideband novelty detection," arXiv preprint arXiv:2505.05903, 2025.

[25] D. Gao, X. Zeng, J. Wang, and Y. Su, "Application of LSTM network to improve indoor positioning accuracy," Sensors, vol. 20, no. 20, p. 5824, 2020.

[26] Y. Tian, Z. Lian, P. Wang, M. Wang, Z. Yue, and H. Chai, "Application of a long short-term memory neural network algorithm fused with Kalman filter in UWB indoor positioning," Scientific Reports, vol. 14, 01 2024.

[27] M. Ren, J. Wei, J. Qin, X. Guo, H. Wang, and S. Li, "Attention based lstm framework for robust UWB and INS integration in nlos environments," Scientific Reports, vol. 15, 07 2025.

[28] Decawave Ltd., DW1000 User Manual, Decawave Ltd., 2017, version 2.18.

[29] TDK InvenSense, ICM-20948: World's Lowest Power 9-Axis MEMS MotionTracking Device, TDK Corporation, Jun. 2017, document No. DS-000189, Revision 1.3.

[30] J. Solà, "Quaternion kinematics for the error-state kalman filter," CoRR, vol. abs/1711.02508, 2017.

[31] S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.

[32] F. Gers, J. Schmidhuber, and F. Cummins, "Learning to forget: continual prediction with LSTM," in 1999 Ninth International Conference on Artificial Neural Networks ICANN 99. (Conf. Publ. No. 470), vol. 2, 1999, pp. 850–855.