Foundation Model-Aided Hierarchical Control for Robust RIS-Assisted Near-Field Communications

TL;DR

This work tackles RIS-assisted near-field communications (NFC) in 6G by addressing the dual-timescale control problem: rapid millisecond-scale CSI updates and slower second-scale blockage dynamics. It introduces a Dual-Transformer HDRL (DT-HDRL) framework that couples a transformer-based NFC CSI estimator with a ViT-based blockage predictor inside a two-timescale hierarchical reinforcement learning controller, enabling proactive path selection (LoS vs RIS) and real-time beamfocusing and phase-shifting. Key contributions include the design of two specialized transformers, the integration into meta- and sub-controllers, and validation on the DeepVerse 6G dataset showing an approximately 18% spectral-efficiency gain and a blockage-prediction F1-score of 0.92 with about 769 ms lead time. The approach demonstrates real-time feasibility on GPU hardware and offers practical guidance for scalable RIS dimensioning in NFC, highlighting the benefit of aligning learning architectures with NFC physics and environmental dynamics.

Abstract

The deployment of extremely large aperture arrays (ELAAs) in sixth-generation (6G) networks could shift communication into the near-field communication (NFC) regime. In this regime, signals exhibit spherical wave propagation, unlike the planar waves in conventional far-field systems. Reconfigurable intelligent surfaces (RISs) can dynamically adjust phase shifts to support NFC beamfocusing, concentrating signal energy at specific spatial coordinates. However, effective RIS utilization depends on both rapid channel state information (CSI) estimation and proactive blockage mitigation, which occur on inherently different timescales. CSI varies at millisecond intervals due to small-scale fading, while blockage events evolve over seconds, posing challenges for conventional single-level control algorithms. To address this issue, we propose a dual-transformer (DT) hierarchical framework that integrates two specialized transformer models within a hierarchical deep reinforcement learning (HDRL) architecture, referred to as the DT-HDRL framework. A fast-timescale transformer processes ray-tracing data for rapid CSI estimation, while a vision transformer (ViT) analyzes visual data to predict impending blockages. In HDRL, the high-level controller selects line-of-sight (LoS) or RIS-assisted non-line-of-sight (NLoS) transmission paths and sets goals, while the low-level controller optimizes base station (BS) beamfocusing and RIS phase shifts using instantaneous CSI. This dual-timescale coordination maximizes spectral efficiency (SE) while ensuring robust performance under dynamic conditions. Simulation results demonstrate that our approach improves SE by approximately 18% compared to single-timescale baselines, while the proposed blockage predictor achieves an F1-score of 0.92, providing a 769 ms advance warning window in dynamic scenarios.

Figures (8)

• Figure 1: System architecture showing dual transformer inputs (CSI transformer from ray-tracing simulations and ViT from camera sensors) feeding into HDRL with high-level strategic planning and low-level beamfocusing optimization.
• Figure 2: Training and validation loss comparison. The transformer achieves 73% lower final MSE with stable convergence.
• Figure 3: Normalized MSE for different received SNR. The CSI transformer maintains low estimation error even in noisy conditions.
• Figure 4: Training loss convergence for blockage prediction. ViT achieves lower final loss with slightly faster initial convergence.
• Figure 5: Average accumulative reward (sum SE) for different simulation steps.