AI/ML
Near-Field Beam Tracking
Traditional near-field beam tracking methods relying on mobility models are fatal in ultra-massive MIMO systems, where even the slightest error could result in beam tracking failures. Thus, the proposed near-field tracking aims to maintain a stable beamforming gain by tracking the mobile station through the
analysis of received signals without requiring mobile dynamics.
By utilizing deep Q-network, the proposed algorithm strengthens its tracking capability from online experiences and updates the combining beam towards positions expected to maximize beamforming gain. Throughout simulations, we compare the proposed algorithm with the Bayesian filter-based methods and confirmed the robustness of the proposed method, especially for abrupt changes in mobile dynamics.
Fig. 1. Procedure for the proposed near-field tracking
Video 1. Near-field tracking
[Related publication]
Radio SLAM
In radio SLAM, reducing information loss and computational complexity is a critical challenge to enable accurate and real-time mapping. As shown in Fig. 1, we propose a machine learning-driven landmark mapping algorithm using channel impulse responses (CIRs). The algorithm leverages CIRs, which preserve the raw characteristics of radio propagation, to enhance mapping performance. Furthermore, it replaces mapping filters in existing radio SLAM algorithms with machine learning to reduce computational complexity and improve efficiency.
Fig. 1. Example of machine learning-driven landmark mapping using CIRs
Video 1. Near-field radio SLAM_2D
Video 2. Near-field radio SLAM_3D
[Related publication]
Localization
Wireless localization determines the position of devices by analyzing characteristics of radio signals such as time of arrival, signal strength, and angle of arrival. It is technically important because accurate location information is essential for high-precision positioning in 6G wireless communication systems. Our lab has explored a wide range of localization methods, including model-based and AI-driven approaches, e.g. cooperative localization and fingerprinting [1, 2]. Recently, we have focused on graph neural networks (GNNs), which provide a powerful way to model the relationships between base stations and user devices using a graph structure [3]. This helps the model better understand the wireless network and improve localization accuracy, especially in complex environments. To validate our theories in practice, we perform real-world prototyping using Ultra-Wideband (UWB) hardware, including demonstrations with small autonomous vehicles for cooperative localization and multi-agent scenarios involving drone swarms and ground rovers.
Fig 1. Illustration of graph-based modeling in wireless communication scenario
Video 1. ADMM based vehicle cooperative positioning POC using V2X test bed