Adaptive Robot Localization with Ultra-wideband Novelty Detection

TL;DR

This work tackles indoor robot localization with Ultra-Wideband by addressing environment-induced range unreliability through novelty detection. It combines a semi-supervised overcomplete autoencoder (NDNN) to quantify per-anchor range novelty with an Extended Kalman Filter that dynamically maps novelty scores to covariance and bias terms, yielding a Nov-EKF that adapts to both static and time-varying disturbances. The approach is trained on nominal data and validated across nine challenging indoor scenarios, demonstrating substantial reductions in positioning error compared with a static EKF and showing robustness to varying numbers of anchors. The results indicate that integrating novelty-guided uncertainty with classical filtering offers a practical, scalable solution for reliable indoor localization in cluttered environments.

Abstract

Ultra-wideband (UWB) technology has shown remarkable potential as a low-cost general solution for robot localization. However, limitations of the UWB signal for precise positioning arise from the disturbances caused by the environment itself, due to reflectance, multi-path effect, and Non-Line-of-Sight (NLOS) conditions. This problem is emphasized in cluttered indoor spaces where service robotic platforms usually operate. Both model-based and learning-based methods are currently under investigation to precisely predict the UWB error patterns. Despite the great capability in approximating strong non-linearity, learning-based methods often do not consider environmental factors and require data collection and re-training for unseen data distributions, making them not practically feasible on a large scale. The goal of this research is to develop a robust and adaptive UWB localization method for indoor confined spaces. A novelty detection technique is used to recognize outlier conditions from nominal UWB range data with a semi-supervised autoencoder. Then, the obtained novelty scores are combined with an Extended Kalman filter, leveraging a dynamic estimation of covariance and bias error for each range measurement received from the UWB anchors. The resulting solution is a compact, flexible, and robust system which enables the localization system to adapt the trustworthiness of UWB data spatially and temporally in the environment. The extensive experimentation conducted with a real robot in a wide range of testing scenarios demonstrates the advantages and benefits of the proposed solution in indoor cluttered spaces presenting NLoS conditions, reaching an average improvement of almost 60% and greater than 25cm of absolute positioning error.

Figures (11)

• Figure 1: The complete pipeline of the proposed adaptive UWB positioning solution. UWB ranges are used to feed a novelty autoencoder, the novelty score is then dynamically mapped into a correction bias and a covariance value for the correction step of the EKF.
• Figure 2: The autoencoder architecture.
• Figure 3: Mapping functions used to transform the estimated novelty score into a bias (right) and a covariance (left) term for the UWB correction step of the Extended Kalman Filter. These functions are kept equal for all the experimental scenarios. The bias term is estimated from novelty scores using a mapping function able to refine the range value passed to the EKF.
• Figure 4: The laboratory equipped with a Vicon tracking system and the Jackal rover used for the experiments. Obstacles are placed in different positions around the area to create the testing scenarios.
• Figure 5: Testing scenarios designed for this paper. The first three scenarios represent distinct trajectories conducted under LoS conditions.