PaperlayerView exact revision on arXiv

Source-linked AI summary

BeamGuard: Risk-Aware Multimodal Beam Forecasting and Adaptive Virtual Beamwidth Control for 6G mmWave V2I Links

Abidemi Orimogunje, Dejan Vukobratovic, Sunwoo Kim, Igbafe Orikumhi, Vukan Ninkovic, Evariste Twahirwa, Gaspard Gashema

arXiv:2608.25433v1eess.SP

TL;DR

Reliable mmWave V2I beam management must handle mobility, blockage, and sensing-domain variation without excessive beam-training overhead. BeamGuard forecasts future beam distributions from multimodal sensing and optional masked in-band measurements, then selects risk-aware virtual beamwidth actions. The framework reports strong ranked forecasting and low outage, while its evidence remains bounded by codebook-level virtual coverage and limited deployment settings.

  • Problem

    Beam management must convert uncertain future beam predictions into actions that balance link reliability, gain, switching cost, and beam-training overhead.

  • Method

    BeamGuard fuses camera, radar, LiDAR, GPS, and optional partial or full mmWave power observations, then plans beam center and virtual beamwidth from predicted posteriors.

  • Results

    BeamGuard demonstrates strong ranked beam forecasting, low threshold-based outage, and practical reliability–overhead tradeoffs across DeepSense 6G evaluations.

  • Takeaways & Limitations

    The framework supports sensor-only, partial in-band, and full hybrid beam management through explicit observation masking and risk-aware virtual coverage control.

  • Takeaways & Limitations

    The study evaluates codebook-level virtual beamwidths rather than physically synthesized wide beams and does not establish universal cross-scenario generalization.

Abstract

from arXiv · show

Reliable beam management is a central challenge for 6G millimeter-wave (mmWave) vehicle-to-infrastructure (V2I) links, where narrow beams provide high array gain but are vulnerable to mobility-induced misalignment, blockage, and domain variation. BeamGuard is a multimodal sensing-aided beam-management framework that combines exteroceptive sensing with optional partial in-band mmWave power observations to forecast future beam distributions and select adaptive virtual beamwidth actions for reliable V2I control. It fuses camera, radar, LiDAR, GPS, and mmWave power observations with a temporal multimodal forecaster, then converts the predicted posterior into a beam center and virtual codebook-level beamwidth through a risk-aware planner. Here, virtual beamwidth denotes adjacent-beam coverage in the codebook index space rather than physical analog wide-beam synthesis. BeamGuard supports sensor-only operation for beam-training overhead reduction, limited in-band operation with masked beam-power entries, and full hybrid operation with sensing and communication-side measurements. We evaluate BeamGuard on DeepSense 6G Scenarios 32 and 33, with additional held-out tests on Scenarios 31 and 34, covering day--night training, transfer, limited adaptation, ablations, budget sweeps, and lightweight baselines. The full-hybrid anchor, used as the complete-system reference, achieves Top-1/Top-3/Top-5 accuracies of approximately \(0.393/0.778/0.897\), while the planner attains a threshold-based outage probability of about \(0.0060\) with a gain ratio of about \(0.895\). Matched-budget baselines further show that BeamGuard improves over multilayer perceptron, recurrent, and temporal convolutional predictors under comparable in-band observation settings. These results demonstrate robust, overhead-aware beam management through multimodal forecasting and risk-aware virtual beamwidth control.

I. Introduction

BeamGuard frames mmWave V2I beam management as forecasting future beam distributions and converting uncertainty-aware predictions into adaptive beamwidth actions. It combines multimodal sensing with optional masked in-band measurements to balance reliability, gain, switching, and training overhead.

  • Motivation: Rapid motion, blockage, scattering changes, and day–night domain shifts make timely narrow-beam alignment unreliable.These conditions motivate forecasting beyond the strongest instantaneous beam.
  • Motivation: Conventional sweeping, tracking, and fixed-codebook methods face an overhead–reliability tradeoff between fine search, narrow-beam gain, and wider coverage.Exhaustive search increases training overhead, while narrow fixed beams are fragile under mobility and blockage.
  • Motivation: Beam prediction accuracy alone does not specify how future beam distributions should become actions balancing reliability, retained gain, switching cost, and beam-training overhead.The paper identifies beam control as the practical gap beyond top-K accuracy.
  • Framework: BeamGuard forecasts future beams from camera, radar, LiDAR, GPS, and optional partial or full mmWave power observations, then plans a center and virtual beamwidth.The planner uses estimated outage risk, expected gain, and switching cost.
  • Problem formulation: BeamGuard formulates beam management as mapping future distributions to virtual beamwidth actions accounting jointly for miscoverage risk, gain, and switching cost.The system uses short histories of sensing and optional communication-side measurements under an observation budget.

B. Beam Forecasting, Virtual Coverage, and Beam-miscoverage risk

BeamGuard forecasts future best-beam posteriors and auxiliary power structure, then uses the next-step posterior to choose a centered virtual coverage action. Miscoverage risk is the posterior mass outside that covered codebook region.

  • Beam forecasting: The forecaster predicts future best-beam probabilities over a horizon and an auxiliary power vector for communication-consistent beam structure.Each predicted probability represents the chance that a beam is best at a future time.
  • Virtual coverage: Virtual beamwidth is the number of adjacent codebook beams covered by an action, not a physically synthesized analog wide beam.Odd widths allow symmetric centering around the selected beam.
  • Virtual coverage: Coverage uses non-circular beam-index distance, so edge regions are clipped rather than wrapped in the finite 64-beam field of view.This reflects the receiver’s ordered, non-cyclic codebook.
  • Risk control: The receding-horizon planner executes the first action using the next-step posterior rather than the entire forecast horizon.The action’s covered posterior mass determines its predicted miscoverage risk.
  • Risk control: Predicted miscoverage risk decreases as an action covers more posterior mass, while narrower actions preserve gain but are more sensitive to uncertainty.This risk proxy supports explicit reliability–gain tradeoffs in planning.

C. Risk-aware Beamwidth control problem

BeamGuard formulates beam management as a finite receding-horizon control problem that selects beam centers and virtual widths using forecasted coverage, gain, miscoverage risk, width cost, and switching cost.

  • Candidate action set: BeamGuard forms candidate centers from top-ranked posterior and predicted-power beams, then combines them with allowable virtual beamwidths.Duplicate centers are removed before constructing the finite action set.
  • Risk-aware objective: The control objective rewards predicted coverage and gain while penalizing excessive width, switching, and miscoverage risk.The risk penalty applies to the amount by which predicted miscoverage exceeds the configured budget.
  • Risk screening: BeamGuard uses a conservative risk screen and a fallback rule when no candidate action satisfies the screening threshold.The fallback preserves risk-minimizing behavior while preventing undefined actions.
  • Receding-horizon control: The planner forecasts over multiple future steps but executes only the first action before repeating the process at the next time instant.This receding-horizon formulation evaluates coverage, gain, width, and switching behavior alongside beam prediction.

III. BEAMGUARD: Multimodal Forecasting and Risk-aware Beamwidth control

BeamGuard connects multimodal future-beam forecasting with communication-aware control by encoding active sensing and power modalities, fusing their temporal representations, and producing posterior and auxiliary power forecasts for planning.

  • Framework overview: BeamGuard maps sensing histories and optional mmWave power observations to beam centers and virtual widths through a risk-aware control pipeline.The design explicitly supports sensor-only, partial in-band, and hybrid operation.
  • Multimodal encoding: The forecaster represents each active modality with an encoder before fusing modality embeddings into a temporal token sequence.The fused history-token sequence is processed by the temporal encoder and decoded across the forecasting horizon.
  • Forecast outputs: An autoregressive decoding head outputs future beam logits and auxiliary power predictions, from which beam posteriors are obtained for planning.The auxiliary power task preserves relative codebook power structure beyond the strongest-beam label.
  • Forecast outputs: The auxiliary power forecast retains neighboring-beam information that the strongest-beam label alone does not represent.This supports modeling angular uncertainty around the best beam.

B. Beam-observation masking and Operating Modes

BeamGuard uses masked beam-power inputs and a shared forecasting pipeline to support sensor-only, partial in-band, and full hybrid operating modes under different observation budgets.

  • Operating modes: The system changes its active modality set to switch among sensor-only, in-band-only, and hybrid forecasting modes.Hybrid operation combines exteroceptive sensing with partial or full in-band power observations.
  • Observation masking: The beam-observation mask marks which power-vector entries are valid measurements rather than treating unobserved entries as low-power beams.For partial observation, the power encoder receives both the masked power vector and its binary mask.
  • Observation masking: The same M-dimensional power representation and masking mechanism support multiple operating modes without changing the forecasting architecture.This enables trading beam-training overhead against communication reliability within one model.
  • Training objective: The forecaster is trained with beam classification, auxiliary power, and beam-distance-aware objectives over the prediction horizon.The distance regularizer discourages probability mass on beams far from the measured best beam.
  • Evaluation protocol: Model selection uses validation data, while final performance is reported only on the held-out test split.

D. Calibration and Posterior selection

BeamGuard calibrates forecast posteriors for risk-aware planning, selects candidate actions through bounded enumeration, and favors wider coverage when uncertainty makes greedy narrow-beam selection risky.

  • Posterior calibration: A scalar temperature is fitted on calibration data because the planner uses posterior mass to estimate coverage and miscoverage risk.Calibration minimizes negative log-likelihood before posterior selection.
  • Posterior selection: The calibrated posterior is retained only when validation ECE improves without increasing validation NLL; otherwise, the raw posterior is used.This rule prevents calibration from degrading already well-behaved posteriors.
  • Action evaluation: The planner evaluates candidate centers from posterior and auxiliary-power rankings across allowable widths, computing coverage, miscoverage risk, and predicted gain.
  • Action evaluation: With Kc = 10 and widths {1, 3, 5}, deterministic enumeration evaluates at most 60 center–width actions before duplicate removal.The resulting computation is negligible relative to multimodal feature extraction and temporal forecasting.
  • Risk-aware planning: When the posterior is spread across adjacent beams, wider virtual coverage can reduce miscoverage risk despite slightly lower gain.Narrow beams are preferred for confident forecasts, while wider coverage may be selected under high uncertainty.
  • Receding-horizon execution: The complete procedure repeatedly encodes observations, forecasts future posteriors, calibrates the selected posterior, and executes the first planned action.

IV. Experimental setup, Ablations and Evaluation Metrics

The evaluation uses controlled, segment-level protocols with synchronized multimodal V2I data, fixed forecasting settings, and separate validation, calibration, and test partitions. Experiments assess forecasting and communication-level control across operating conditions, budgets, and risk settings.

  • BeamGuard is evaluated through joint day–night, transfer, adaptation, budget, operating-regime, risk-sensitivity, and lightweight-baseline experiments.
  • The dataset contains synchronized receive-power vectors and standardized sensor records with best-beam labels and modality paths.
  • 6600 synchronized rows are split into 4000 training, 800 validation, 800 calibration, and 1000 test rows under the joint day–night protocol.
  • Validation selects checkpoints and hyperparameters, calibration fits temperature scaling, and the held-out test partition is reserved for final reporting.
  • The standard configuration uses 100 ms sampling, H = 8 history samples, T = 5 future samples, and a fixed candidate-center count Kc = 10.

B. Evaluation protocols

The protocols test in-domain performance, day–night transfer, limited adaptation, held-out scenarios, modality regimes, partial-observation masks, and planner sensitivities under controlled settings.

  • Evaluation protocols: Zero-shot and few-shot protocols evaluate cross-scenario generalization and improvement from limited target-domain supervision without training from scratch.
  • Operating regimes: Operating regimes span sensor-only, power-only, GPS–power, GPS–LiDAR–power, and full camera–radar–LiDAR–GPS–power fusion.
  • Beam-observation budget: Partial beam observation varies measured entries in the M = 64 codebook, with an explicit mask distinguishing unobserved beams from measured low-power beams.
  • Mask sensitivity: Uniform and local-neighborhood probing are compared while model settings, partitions, budgets, planner parameters, and seeds remain fixed.
  • Planner ablations: Risk-budget, fixed-width, and same-forecaster controller ablations isolate conservativeness, adjacent-beam coverage, and controller logic.

D. Baselines

The baseline suite controls data splits, horizons, codebook targets, and calibration to compare persistence, static, recurrent, convolutional, sensing-aided, and transformer-style predictors alongside planner rules.

  • Predictor baselines: All learned predictors share segment-level partitions, H = 8, T = 5, and M = 64 targets, while persistence uses the same test windows.
  • Predictor baselines: Baselines include persistence, power-only MLP, GRU, LSTM, temporal CNN, sensor-only temporal CNN, and BeamGuard’s transformer-style forecaster.
  • Sensing-aided references: C+G and Lid+G provide controlled vision-position and LiDAR-position references rather than exact reimplementations of published methods.
  • Controller baselines: Planner comparisons include greedy fixed-width rules, a widest-safe heuristic, a risk-ablated adaptive controller, and the complete risk-aware planner.
  • Metrics: TopKacc measures ranked beam inclusion, while NLL, Brier score, ECE, DBA@3, power metrics, outage, gain ratio, and switching rate assess predictive and control behavior.

F. Reproducibility and Code Availability

BeamGuard’s results connect multimodal future-beam forecasting with risk-aware virtual beamwidth control and examine reliability, gain, overhead, transfer, and observation-budget behavior.

  • Anchor results: The full-hybrid anchor achieves Top-1, Top-3, and Top-5 accuracies of approximately 0.393, 0.778, and 0.897.
  • Operating regimes: The Full regime has the lowest threshold-based empirical outage in the operating comparison, despite not uniformly maximizing Top-Kacc accuracy.
  • Beam-observation budget: Increasing observed entries generally improves ranked prediction and reduces outage, while partial budgets retain much of full-budget behavior.
  • Beam-observation budget: The Brier score decreases from 0.7465 at L = 8 to 0.7320 at L = 64, while DBA@3 increases from 0.9046 to 0.9149.
  • Mask sensitivity: Uniform and local probing produce comparable accuracy, outage, gain-ratio, and switching results, with neither policy uniformly dominating.

D. Day-Night Transfer and Few-shot Adaptation

Day–night transfer is harder than in-domain evaluation, but hybrid sensing and limited target-domain adaptation improve BeamGuard’s transfer behavior. Better ranked forecasting does not uniformly translate into lower outage.

  • Zero-shot transfer: Hybrid configurations remain competitive under zero-shot day–night transfer by combining complementary sensing and communication evidence.G+Lid+Pwr combines coarse mobility, geometric structure, and in-band power measurements, while full hybrid uses the complete sensing stack.
  • Few-shot adaptation: 20% target-domain adaptation consistently improves ranked beam forecasting relative to zero-shot transfer.For S32→S33 full hybrid, Top-1/Top-3/Top-5 rises from 0.1403/0.3184/0.5096 to 0.2779/0.5890/0.7376; for S33→S32, it rises from 0.1933/0.4252/0.5436 to 0.3108/0.6654/0.8145.
  • Few-shot adaptation: The G+Lid+Pwr configuration shows the same ranked-prediction improvement trend under limited target-domain supervision.This indicates that the transfer benefit does not require the full sensing stack.
  • Few-shot adaptation: Planner-level outcomes improve in most transfer settings, especially for S33→S32, but improved ranked prediction does not always guarantee lower threshold-based outage.The result supports few-shot adaptation while preserving a clear distinction between forecasting and communication-control metrics.

E. Additional Held-out Scenario Generalization

Held-out Scenarios 31 and 34 impose stronger domain shifts than the original S32/S33 day–night evaluation. Communication-side observations improve reliability in these tests, while adaptive risk-aware control balances outage, gain, and switching behavior.

  • Held-out scenario transfer: Held-out scenario transfer is substantially more challenging than the S32/S33 day–night evaluation.Ranked prediction accuracies decrease sharply on both S31 and S34, consistent with stronger shifts in road geometry and scenario distribution.
  • Held-out scenario transfer: On S31, G+Pwr gives the lowest threshold-based outage and highest gain ratio among the evaluated regimes.On S34, full hybrid gives the best ranked prediction, lowest outage, and highest gain ratio among the evaluated regimes.
  • Risk-budget sensitivity: For full hybrid, Pout(0.40) remains between 0.0057 and 0.0067 while Rgain remains between 0.8975 and 0.9082 across the tested risk budgets.From β = 0.05 to 0.15, switching rate falls from 0.1622 to 0.1298, a 20.0% reduction.
  • Risk-budget sensitivity: Increasing β reduces switching consistently while outage and gain vary only marginally and non-monotonically.The observed effect is a reliability–temporal-stability tradeoff rather than a pronounced outage–gain tradeoff.
  • Planner behavior: The planner uses the predicted beam distribution and auxiliary power estimate to select beam center and virtual beamwidth while accounting for reliability, gain, width, and switching costs.The fixed-width ablation isolates adjacent-beam coverage, and adaptive BeamGuard jointly selects center and width from the same forecast inputs.
  • Planner behavior: Controller logic changes communication-level behavior independently of predictor accuracy.The risk-aware planner reduces switching by 13.8% relative to the no-risk controller and approximately 43.2% relative to greedy fixed-width controllers, with lower covered-power gain ratio.

H. Predictor Baselines

BeamGuard’s predictor baselines distinguish matched-input forecasting strength from complete-system beam-management performance, while runtime and deployment analyses clarify practical tradeoffs and remaining validation needs.

  • Matched predictor baselines: BeamGuard G+Pwr provides the strongest performance in the matched G+Pwr, L = 64 predictor comparison.All competing models use the same GPS and full in-band power inputs, isolating forecasting architecture from modality and observation-budget differences.
  • Complete-system evaluation: The full-hybrid configuration is evaluated separately as the complete-system reference, combining predictor results with outage, gain, switching, and latency metrics.Table X instead evaluates predictors under controlled input settings.
  • Runtime and complexity: 39.56 million parameters and 151.17 MB characterize the full-hybrid anchor, with 76.37 ms predictor inference and 1.49 ms planner overhead per window.Profiling used batch size 1 on an Apple M2 Pro with PyTorch MPS acceleration.
  • Runtime and complexity: The predictor accounts for approximately 98.1% of measured latency, while the risk-aware planner contributes about 1.9%.The planner is therefore a small fraction of the measured software pipeline latency, although the measurements are platform-specific.
  • Deployment tradeoffs: G+Pwr is an attractive lower-complexity operating point when communication-side power measurements are available, whereas full fusion provides richer environmental context for difficult conditions.G+Lid+Pwr offers an intermediate configuration with fewer sensing branches than the full-hybrid model.
  • Key findings and scope: The evaluation supports coupling multimodal forecasting with explicit virtual beamwidth control, but embedded hardware validation remains necessary before deployment-level latency and energy claims.Held-out S31/S34 tests also expose difficulty with broader scenario transfer, and the study uses virtual rather than physically synthesized wide beams.
Loading 2608.25433v1…