PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Modal Environment-Aware Beam Management for Massive MIMO: A Geometry-Driven Virtual Base Station Framework

Yijie Bian, Wei Guo, Jie Yang, Shenghui Song, Jun Zhang, Shi Jin, Khaled B. Letaief

arXiv:2606.26567v1eess.SPcs.AI

TL;DR

High-frequency MU-MIMO beam management faces costly beam training and difficult interference-aware coordination. This paper builds a geometry-driven VBS database from LiDAR and location data, then uses coarse reconstruction, partial training, and hierarchical reinforcement learning for beam selection. Simulations on realistic urban layouts show improved training efficiency and beam-selection performance over baseline methods.

  • Problem

    High-frequency MU-MIMO beam management must acquire costly CSI and coordinate combinatorial joint beam choices, while sensing-aided approaches have limited coverage of NLoS paths and multi-user interference.

    Source: A. Prior Works and Limitations · p. 2, ; NUE · p. 3

  • Method

    The framework derives an interpretable VBS database from environmental geometry, uses it for coarse channel reconstruction and VOP partial training, and applies DD3QN-CBS for coordinated beam selection.

    Source: B. Main Contributions · p. 2, 2) Path Loss Computation · p. 6, C. VBS-guided DD3QN-CBS · pp. 7–8

  • Results

    Simulations on realistic urban layouts show improved beam-training efficiency and beam-selection performance over baseline methods.

    Source: §VI · p. 13

  • Takeaways & Limitations

    The results support geometry-aware VBS representations as an interpretable interface for representing dominant LoS and VLoS opportunities in beam management.

    Source: §VI · p. 13

    Existing sensing-aided approaches can miss dominant urban NLoS paths when they rely only on LoS information, and reported multi-user approaches include LoS-only or sparse-user settings.

    Source: A. Prior Works and Limitations · p. 2

Abstract

from arXiv · show

High-frequency massive multiple-input multiple-output (MIMO) systems promise ultra-high data rates. However, efficient beam management remains challenging due to the prohibitive beam training overhead and intricate coordination required in multi-user MIMO (MU-MIMO) scenarios. To address these bottlenecks, environment-aware communications have emerged as a promising paradigm, leveraging site-specific knowledge to circumvent exhaustive pilot-based beam training and streamline multi-user communications. In this paper, we propose an interpretable and geometry-driven framework that utilizes multi-modal environmental data, specifically regional 3D light detection and ranging (LiDAR) point clouds and location information, to construct an offline virtual base station (VBS) database. By modeling dominant reflection paths via mirror symmetry across building facades reconstructed from the point clouds, the VBS database provides a compact and sparse description of the wireless propagation environment. To bridge the semantic gap between geometric information and wireless channels, we develop a coarse channel reconstruction mechanism that estimates channel parameters directly from VBS-derived geometric relationships. Based on the resulting coarse beamspace representation, we design a VBS-assisted orthogonal-pilot (VOP)-based partial beam training scheme to refine the coarse estimates with minimal online training overhead. Finally, to tackle the combinatorial beam selection problem and manage inter-user interference, we propose a hierarchical deep reinforcement learning framework, namely a dual-agent dueling double deep Q-network, for coordinated beam selection (DD3QN-CBS). Simulation results demonstrate consistent gains in both beam training efficiency and beam selection performance over heuristic and learning-based baselines.

I. INTRODUCTION

High-frequency MU-MIMO beam management is constrained by costly CSI acquisition and combinatorial, interference-aware beam coordination. The paper proposes an interpretable geometry-driven VBS framework that uses environmental data to guide beam training and coordinated selection.

I. INTRODUCTION

  • Exhaustive training and real-time high-dimensional CSI acquisition impose substantial overhead, while coordinating beam pairs across users creates a combinatorial search under interference.These constraints are especially challenging with hybrid architectures that have far fewer RF chains than antenna elements.Source: §I · p. 1, ; NUE · p. 3

A. Prior Works and Limitations

  • Prior sensing-aided work often targets single-user or LoS settings, while multi-user approaches can leave interference unmodeled or entangle sensing with coordination.The cited limitations include evaluations with sparse user distributions and high computation from processing raw multi-modal data.Source: A. Prior Works and Limitations · pp. 1–2

B. Main Contributions

  • The framework converts LiDAR point clouds and BS location into a sparse, interpretable VBS propagation representation for interference-aware MU-MIMO beam management.VBSs are virtual mirror images across reflective facades and encode site-specific propagation structures.Source: B. Main Contributions · p. 2

VBS database construction from multi-modal data.

  • The VBS database combines VBS locations with LoS and VLoS coverage regions to compactly describe dominant propagation paths without preliminary wireless measurements.The paper presents this as a site-specific representation with low storage overhead.Source: B. Main Contributions · p. 2

B. Main Contributions

  • Because the VBS database provides only a coarse prior, VOP-based partial beam training uses reconstructed channel parameters to calibrate residual geometric uncertainty and propagation variation.The authors state that this scheme reduces beam-training overhead.Source: B. Main Contributions · p. 2

II. SYSTEM MODEL AND FRAMEWORK MOTIVATION

  • The paper frames VBS abstraction as a geometry-driven link between high-frequency propagation characteristics and the subsequent beam-management pipeline.Its system-level foundation first formulates downlink MU-MIMO beam selection, then motivates the abstraction.Source: §II · p. 2, ; NUE · p. 3

A. Beam Selection Problem Formulation

  • The downlink system model considers a hybrid-beamforming BS serving co-scheduled users, with analog beam choices drawn from discrete BS and user codebooks.The beam-selection objective maximizes system ESE subject to codebook and total-transmit-power constraints.Source: A. Beam Selection Problem Formulation · pp. 2–3, ; NUE · p. 3

B. Concept Generation and Formulation of VBS

The VBS abstraction represents facade reflections as virtual BS locations, providing a compact geometric description of propagation opportunities. The database stores those locations with quantized coverage regions for environment-aware beam management.

B. Concept Generation and Formulation of VBS

  • At high frequencies, attenuation and blockage can make direct LoS unreliable, while static building facades can provide persistent auxiliary reflection paths.This motivates representing reflected links geometrically.Source: B. Concept Generation and Formulation… · p. 3
  • For a reflecting facade, a VBS is the mirror image of the physical BS and represents a single-bounce path through that reflector.The paper distinguishes this beam-management role from virtual anchors used as geometric beacons for positioning.Source: B. Concept Generation and Formulation… · p. 3
  • Facade-induced virtual links are treated as structurally permanent, unlike ground reflections that are susceptible to dynamic moving occlusions.The paper uses this distinction to motivate a compact geometric basis for MU-MIMO beam selection.Source: B. Concept Generation and Formulation… · p. 3
  • The VBS database stores virtual-station locations and quantized coverage regions to make the geometric abstraction usable for beam management.It associates each VBS with a VLoS coverage region and the physical BS with the LoS region.Source: B. Concept Generation and Formulation… · p. 3
  • Together, the physical BS, VBS locations, and coverage regions encode path availability and angular priors for beam selection.Stable VLoS paths can provide auxiliary links that spatially separate co-scheduled users and mitigate inter-user interference.Source: B. Concept Generation and Formulation… · p. 3
  • The offline database-construction procedure proceeds through LiDAR acquisition, geometric extraction, VBS location computation, and coverage computation and storage.The procedure is described as requiring no wireless measurements.Source: §III · p. 3

A. Geometric Information Extraction

The geometric extraction pipeline turns static LiDAR structures into reflector meshes, then uses geometric validity checks and clustering to produce compact VBS locations. The resulting abstraction represents dominant single-bounce VLoS opportunities when LoS is blocked.

A. Geometric Information Extraction

  • Static point clouds are segmented into building objects and coherent clusters, whose triangular meshes approximate candidate reflective surfaces.The described extraction uses DBSCAN before mesh generation and surface simplification.Source: A. Geometric Information Extraction · pp. 3–4
  • Facade reflections can provide dominant VLoS paths when LoS is blocked, extending coverage and supplying spatially distinct links for MU-MIMO selection.The figure also depicts VBS locations and quantized coverage regions as a compact abstraction that can replace the original point clouds.Source: Figure 1 · p. 4
  • The method treats constituent facade meshes as independent reflectors and applies mirror geometry to obtain VBS locations from valid facades.The single-bounce assumption is motivated by dominant NLoS components and the typically negligible energy of multi-bounce paths.Source: A. Geometric Information Extraction · pp. 3–4

B. Determination of VBS Locations

  • Candidate VBSs are retained only for meshes satisfying both an orientation condition and an unobstructed-path condition.The orientation check requires the mesh to face the BS; the visibility check accepts a mesh when at least one vertex has an unoccluded incident path.Source: B. Determination of VBS Locations · p. 4
  • Valid VBS locations are computed as mirror images of the BS across mesh planes, with the mesh-plane distance used in the location calculation.The passage defines the resulting VBS as the mirror image of the BS location.Source: B. Determination of VBS Locations · p. 4
  • HDBSCAN clusters nearby raw VBS coordinates into centroid representatives, reducing redundancy from LiDAR noise and surface-reconstruction deviations.Its adaptive neighborhood radius is described as more robust than fixed-radius DBSCAN for facades with varying sizes and orientations.Source: B. Determination of VBS Locations · p. 4

C. Coverage Computation and Database Storage

The offline database stores physical and virtual base-station locations with discretized LoS and VLoS coverage, then supplies geometric path information for coarse channel reconstruction. The resulting representation supports later beam management without retaining raw LiDAR in the online system.

C. Coverage Computation and Database Storage

  • Coverage is represented on a user-height grid: LoS cells have an unobstructed BS path, while VLoS cells intersect the associated reflector and avoid other meshes.The VLoS path must remain unobstructed on both incident and reflected sub-segments.Source: C. Coverage Computation and Database… · pp. 4–5
  • The database stores BS/VBS locations with LoS/VLoS grid coverage, providing a compact alternative to dense channel tables and allowing raw LiDAR to be discarded online [28].Its construction workflow uses regional LiDAR and BS location, building meshes, clustered VBS positions, and quantized coverage areas.Source: C. Coverage Computation and Database… · p. 5, Figure 2 · p. 5

IV. VBS-GUIDED MU-MIMO BEAM MANAGEMENT

  • The offline VBS database is converted into a coarse beamspace representation, which online partial training calibrates before interference-aware multi-user beam selection.These are the framework's sequential stages: coarse reconstruction, partial training, and coordinated selection.Source: §IV · p. 5

A. VBS-assisted Coarse Channel Reconstruction

2) Path Loss Computation

The coarse channel model uses free-space path loss with an added effective reflection loss for VLoS paths because the database lacks material properties. It retrieves nearby coverage-grid paths and aggregates their spatial signatures to rank beam pairs and narrow the search.

2) Path Loss Computation

  • LoS and VLoS path losses use free-space propagation, with an effective reflection-loss term added for VLoS paths to model unknown material attenuation.This simplified model supports beam ranking and search-space narrowing rather than high-fidelity electromagnetic simulation.Source: 2) Path Loss Computation · p. 6
  • Because environmental geometry lacks reliable reflection coefficients and surface attenuation, the coarse reconstruction uses a simplified propagation model rather than high-fidelity electromagnetic simulation.The model is intended to rank candidate beams and narrow the search space.Source: 2) Path Loss Computation · p. 6

3) Coarse Channel Reconstruction

  • To reduce sensitivity to point-cloud noise and grid quantization, the method retrieves K-nearest-neighbor grid sets for LoS and VLoS paths, then aggregates identified paths into a coarse channel preserving spatial signatures.The LoS and reflected-path neighborhoods can have different sizes to reflect different uncertainty levels.Source: 3) Coarse Channel Reconstruction · p. 6

B. VOP-based Partial Beam Training

VOP-based partial training refines geometry-derived coarse channels by measuring promising beam directions online. Separate measured search subsets and broader feasible candidate subsets balance training overhead against retaining joint beam choices.

B. VOP-based Partial Beam Training

  • VOP-based partial training measures only promising directions identified from coarse reconstruction, calibrating estimates affected by geometric and propagation uncertainty.The coarse channel supplies the measurement plan; pilot responses provide channel-aware calibration.Source: B. VOP-based Partial Beam Training · p. 6
  • Two-level beam subsets use smaller search sets for online measurements and broader candidate sets for final coordinated selection.The distinction controls training overhead while retaining feasible joint beam configurations.Source: B. VOP-based Partial Beam Training · p. 6, ; NUE · p. 7

1) Beam Search and Candidate Subset Construction

  • Coarse channels are projected into beam codebooks to form gain matrices, and a greedy highest-magnitude beam-pair procedure orders BS-side and user-side indices for subset construction.The longer prefixes form candidate subsets, while shorter prefixes form the search subsets measured online.Source: 1) Beam Search and Candidate Subset… · p. 6, ; NUE · pp. 6–7

2) Partial Beam Measurements and Feature Synthesis

VOP reduces training by measuring selected beamspace entries and letting users transmit distinguishable pilots concurrently. The resulting online measurements are combined with coarse beamspace features for coordinated selection.

2) Partial Beam Measurements and Feature Synthesis

  • VOP requires ⌈¯NBS/NRF⌉¯NUE training slots, combining reduced beam dimensions with simultaneous distinguishable user pilots instead of sequential user training.Exhaustive sequential training requires ⌈NBS/NRF⌉NUEK slots, while blind full-codebook orthogonal-pilot training requires ⌈NBS/NRF⌉NUE slots.Source: 2) Partial Beam Measurements and… · p. 7
  • All users transmit orthogonal uplink pilots concurrently, and the BS correlates received signals with each pilot to suppress inter-user interference and decode per-user measurements.The pilot length satisfies τ ≥ K.Source: 2) Partial Beam Measurements and… · p. 7, RF )HHT · p. 7
  • Measured beam entries populate online beamspace matrices, whose candidate-subset features are concatenated with coarse beamspace features for a calibrated representation.The two feature sources combine VBS-assisted estimates with online measurements.Source: RF )HHT · p. 7

C. VBS-guided DD3QN-CBS

After candidate pruning, the paper formulates coordinated beam selection as a sequential leader-follower decision problem and uses dual-agent DD3QN-CBS to select beams under inter-user interference. The BS agent chooses first, and the user-side agent conditions its choice on the BS decision, with both agents trained toward the ESE objective.

C. VBS-guided DD3QN-CBS

  • After pruning, joint beam selection remains nontrivial because users’ beam choices are coupled by interference and ESE is available only after joint beam and precoder decisions.Exhaustive search also grows combinatorially across BS and user candidate beams.Source: C. VBS-guided DD3QN-CBS · p. 7
  • The method casts selection as a finite-horizon leader-follower Markov game: the BS chooses a transmit beam, then the user agent chooses a receive beam conditioned on it.It assumes K = NRF simultaneously served users.Source: C. VBS-guided DD3QN-CBS · p. 8
  • The state records beamspace features and prior decisions; action masking prevents invalid beam reuse, while dense intermediate rewards address sparse terminal feedback.The shared terminal objective is maximizing ESE.Source: C. VBS-guided DD3QN-CBS · p. 8
  • Both agents approximate their action-value functions with dueling double deep Q-networks, using double-Q action selection and evaluation to mitigate overestimation bias.The BS evaluates its transmit-beam choice; the user-side value function also conditions on the BS action.Source: C. VBS-guided DD3QN-CBS · p. 8

1) State Space

The state representation combines beamspace information with records of beam choices and online measurements. The action design follows the BS-then-user decision order and masks invalid BS beam reuse.

1) State Space

  • The composite input tensor combines beamspace features with action and measurement information that records selection history and VOP calibration.The action mask tracks beam occupancy; the measurement mask records which entries have been measured.Source: 1) State Space · pp. 8–9

2) Action Space

  • Assigning multiple users the same BS codeword can make the effective channel matrix low-rank and weaken spatial multiplexing, motivating constraints on correlated assignments.The paper uses action masking to focus agents on ranking valid beam pairs rather than penalty-based collision avoidance.Source: 2) Action Space · p. 9
  • The BS-side feasible action set excludes previously selected transmit beams, while a shared user-side Q-network represents K homogeneous user agents.Parameter sharing is described as improving sample efficiency and generalization without changing MU-MIMO coupling.Source: 2) Action Space · p. 9

3) Reward Function

The reward design supplies intermediate feedback in the leader-follower decision order while preserving actual ESE as the terminal objective. The selected analog beams define an effective channel used for digital precoding.

3) Reward Function

  • Intermediate BS- and user-side rewards use estimated SINR proxies, while terminal feedback remains the actual ESE.This is intended to avoid delaying the learning signal until episode end.Source: 3) Reward Function · pp. 9–10
  • The BS-side utility evaluates multi-user coverage while avoiding overlap with selected BS beams; the user-side utility evaluates receive-beam quality conditioned on the BS action and SINR.The resulting rewards are normalized to [0, 1].Source: 3) Reward Function · p. 10
  • Selected BS and user beams form analog precoding and combining matrices, producing an effective channel for MMSE digital precoding.The regularization parameter η is set to N0W/PT to balance noise enhancement and interference suppression.Source: 3) Reward Function · p. 10

4) Network Architecture

The evaluation reports coarse-beamspace reconstruction results in a noisy LiDAR-based urban digital twin, alongside the paper’s compact two-agent network design. The VBS-assisted representation captures dominant directions better than location-only geometry when strong NLoS paths dominate.

4) Network Architecture

  • Both agents use a two-layer CNN, attention module, and fully connected projection to process structured beamspace observations into action representations.The BS network preserves candidate-wise structure; the user-side network conditions on accumulated BS selections.Source: 4) Network Architecture · p. 10

A. Simulation Setup

  • Evaluation uses a Blender–Sionna digital twin of a 120 m × 120 m Mexia urban area, with LiDAR perturbed by 10 cm Gaussian noise and 10% random point drops.The layout is reconstructed from OpenStreetMap data; users are randomly distributed in unoccupied regions.Source: A. Simulation Setup · pp. 10–11

B. Performance of Coarse Channel Reconstruction

  • −3.74 dB and −0.18 dB were the averaged NMSE values for the two cases; VBS-assisted beamspace captured dominant directions better than location-only geometry under strong NLoS.The two cases respectively involved blocked LoS and one LoS path with several auxiliary NLoS paths.Source: B. Performance of Coarse Channel… · p. 11, Figure 4 · p. 11

C. Performance of VOP-based Partial Beam Training

The framework combines VOP-based online calibration with DD3QN-CBS for coordinated beam selection. Results report improved training efficiency and beam-selection performance, with DD3QN-CBS gains across average and lower-tail ESE evaluations.

C. Performance of VOP-based Partial Beam Training

  • VOP achieves low channel-reconstruction NMSE with a small training budget, and accuracy improves as the budget increases.Increasing the user-side measurement budget helps more than increasing the BS-side budget, with gains eventually plateauing.Source: C. Performance of VOP-based Partial… · p. 11

D. Performance of DD3QN-CBS

  • The evaluation compares DD3QN-CBS with stochastic, heuristic, and learning-based beam-selection baselines across RF-chain, noise, and SINR-threshold settings.The baselines include random assignment, VBS-only and measurement-updated maximum-magnitude assignment, and DRL-CBS.Source: D. Performance of DD3QN-CBS · p. 11

1) Convergence Performance

  • DD3QN-CBS training ESE rises faster and remains higher across configurations than the learning-based DRL-CBS baseline.The authors attribute this to action masking and leader-follower coordination, which prune invalid beam combinations during exploration.Source: 1) Convergence Performance · pp. 11–12

2) Average ESE under Different Configurations

  • DD3QN-CBS consistently outperforms MM-BA across noise and SINR-threshold configurations, with its advantage narrowing under stronger noise.Across RF-chain settings, its gain over MM-BA grows as RF chains increase.Source: 2) Average ESE under Different… · p. 12

3) Distribution of ESE under Different Configurations

  • DD3QN-CBS improves lower-tail test ESE by reducing poorly matched beam assignments, while remaining comparable to MM-BA in the high-ESE region.Together with average ESE results, the CDF analysis indicates improved reliability under difficult multi-user coupling.Source: 3) Distribution of ESE under Different… · p. 12

VI. CONCLUSION

  • Simulations on realistic urban layouts show improved beam-training efficiency and beam-selection performance over baseline methods.The framework retains a physically interpretable representation of dominant LoS and VLoS opportunities.Source: §VI · p. 13
Loading 2606.26567v1…