DiffLoc: Diffusion Model-Based High-Precision Positioning for 6G Networks

TL;DR

DiffLoc tackles the challenge of outdoor ultra-precise UE positioning in 6G by mapping high-dimensional massive-MIMO CSI directly to 2D coordinates using a conditional diffusion model. It introduces a principled score-based fusion across multiple base stations and a consistency-training paradigm that enables few-step, real-time inference while preserving sub-centimeter accuracy in challenging urban deployments. The approach progresses from a plain MLP baseline to a UNet-based score network and finally to a two-step consistency model (DiffLoc-CT), achieving 0.5 cm fusion accuracy and 1–2 cm single-BS accuracy in Tokyo UMa simulations. This work demonstrates the practicality of diffusion-driven localization for real-time 6G applications and provides a scalable framework for multi-BS fusion in dynamic outdoor environments.

Abstract

This paper introduces a novel framework for high-accuracy outdoor user equipment (UE) positioning that applies a conditional generative diffusion model directly to high-dimensional massive MIMO channel state information (CSI). Traditional fingerprinting methods struggle to scale to large, dynamic outdoor environments and require dense, impractical data surveys. To overcome these limitations, our approach learns a direct mapping from raw uplink Sounding Reference Signal (SRS) fingerprints to continuous geographic coordinates. We demonstrate that our DiffLoc framework achieves unprecedented sub-centimeter precision, with our best model (DiffLoc-CT) delivering 0.5 cm fusion accuracy and 1-2 cm single base station (BS) accuracy in a realistic, ray-traced Tokyo urban macro-cell environment. This represents an order-of-magnitude improvement over existing methods, including supervised regression approaches (over 10 m error) and grid-based fusion (3 m error). Our consistency training approach reduces inference time from 200 steps to just 2 steps while maintaining exceptional accuracy even for high-speed users (15-25 m/s) and unseen user trajectories, demonstrating the practical feasibility of our framework for real-time 6G applications.

Figures (11)

• Figure 1: Score‐based fusion: each BS $b$ produces a posterior density $p(\mathbf{x}\!\mid\!\mathbf{H}^{(b)})$, shown as a shaded surface. Summing these per-BS score fields yields the fused score $\sum^{B}_{b=1} \nabla_{\mathbf{x}}\log p(\mathbf{x}\!\mid\!\mathbf{H}^{(b)}) = \nabla_{\mathbf{x}}\log p_{\mathrm{fused}}(\mathbf{x}\!\mid\!\mathcal{H})$, which guides the final position estimate (right).
• Figure 2: Channel preprocessing pipeline: reshape $(N_r \times N_t \times N_c)$ CFR into $(N_rN_t)\times N_c$, convert to delay domain via IFFT, then truncate to the first $L=150$ taps.
• Figure 3: DiffLoc score network architectures for conditional UE localization diffusion. (a) DiffLoc-MLP processes concatenated inputs through $[512, 256, 128]$ hidden layers with residual connections ($\sim$16.9M parameters). (b) DiffLoc-UNet employs encoder-decoder structure, progressively expanding features to $[256, 512, 1024]$ then contracting to $[512, 256, 128]$ with skip connections preserving spatial information ($\sim$5.5M parameters). Both networks predict noise vectors $\boldsymbol{\epsilon}$ for reverse diffusion-based UE positioning.
• Figure 4: Layout of the Tokyo urban macro-cell scenario in NVIDIA AODT. Seven BSs (red dots) and UEs (white dots) are shown. Each UE moves at a constant speed to generate the channel dataset.
• Figure 5: Visualization of each method. (a) Ground truth, (b) Supervised MLP, (c) CNN-based method, and (d) Grid-based fusion are baseline methods. (e) DiffLoc-MLP, (f) DiffLoc-UNet, and (g) DiffLoc-CT are our proposed methods. We plot only the final fused results; methods (b) and (c) only have fused plots.