PaperlayerView exact revision on arXiv

Source-linked AI summary

Sensing-Aided Vehicular Beam Tracking: System Gain Over Link Feedback and Probe-Verified Sensing Invocation

Jeongwan Kang, Miguel Bordallo López, Nhan Thanh Nguyen, Junhyeong Kim, Sunwoo Kim

arXiv:2610.20196v2eess.SP

TL;DR

Sensor-based beam prediction is usually compared with beam training, leaving unclear what it adds beyond a deployed tracker’s link feedback and when sensing should be withheld. This paper evaluates a probe-verified controller with separately accounted probes and model invocations in causal replay of recorded vehicular beam sweeps. Sensing’s value depends on feedback limitations and sensing accuracy and timeliness: it helps in the high-innovation scenario when feedback is stressed, but weak or late sensing can be harmful.

  • Problem

    Sensor-based beam predictors are usually compared with exhaustive or hierarchical training rather than feedback-only tracking, leaving their incremental value beyond link feedback unclear.

    Source: §I · p. 1

  • Method

    The paper uses action-restricted causal replay and a controller that admits sensing after probe verification, backs off after failed trials, and performs a post-setup re-sweep.

    Source: §I · pp. 1–2

  • Results

    In three scenarios, feedback-only tracking is within 0.25 dB of the recorded-sweep oracle and always-on sensing is harmful; in the remaining scenario, accurate sensing reduces loss when feedback is stressed.

    Source: §I · p. 2

  • Takeaways & Limitations

    Sensing’s value depends jointly on feedback limitations and the accuracy and timeliness of sensing information.

    Source: §I · p. 2

    Causal replay relies on exhaustive recorded sweeps and idealized within-epoch timing, and re-sweep comparisons in Scenario 9 are test-informed rather than confirmatory.

    Source: §VII · p. 13

Abstract

from arXiv · show

Sensor-based beam prediction is usually evaluated against exhaustive beam training, leaving open when sensing adds useful information beyond a deployed tracker's link feedback and when it should be withheld. We separate system gain from the marginal value of sensing in action-restricted causal replay of recorded 64-beam sweeps from four vehicular DeepSense 6G scenarios. The replay exposes only serving-beam and explicitly probed candidate powers and counts probes and model invocations separately. The proposed controller combines a scheduled post-setup re-sweep with probe-verified sensing admission and blacklist backoff. In three scenarios where feedback-only tracking is within 0.25 dB of the recorded-sweep oracle, the controller stays close to feedback-only in mean loss while rarely invoking the model, whereas always-on sensing performs worse. In the sole high-innovation scenario, the controller's system gain is 0.82 dB (95% confidence interval [0.30, 1.40]) over the plain feedback-only tracker and 0.15 dB over a probe-limited re-sweeping reference, with fewer probes than either. Sensing and the re-sweep both correct setup lock loss. At the native decision interval, enabling sensing in the same controller yields no statistically resolved loss reduction but cuts probing by 21%. In this scenario, accurate sensing reduces loss relative to the sensor-disabled controller under emulated link noise, staleness or longer decision intervals. Weak or delayed sensing can instead increase loss despite probe verification. These exploratory results characterize sensing as a conditional substitute for RF probing or a source of loss reduction, depending on the limitations of link feedback and on the accuracy and timeliness of sensing.

I. INTRODUCTION

The paper asks when sensing adds useful information beyond a deployed link’s feedback and when it should be withheld. It evaluates a probe-verified controller in causal replay and finds sensing’s value depends on both feedback limitations and sensing accuracy and timeliness.

I. INTRODUCTION

  • The central gap is that sensor-based predictors are usually compared with exhaustive or hierarchical training, not a feedback-only tracker using serving-beam and probe measurements.That comparison measures predictor accuracy, but not what sensing adds beyond the link’s existing feedback.Source: §I · p. 1
  • The evaluation restricts policies to acquired link measurements and separately accounts for candidate-beam probes and model invocations, avoiding the optimism of open-loop evaluation.It also identifies oracle re-anchoring as a separate source of optimism.Source: §I · p. 1
  • The controller combines probe-verified sensing admission, blacklist backoff after failed trials, and a post-setup re-sweep.A trial probes the proposed beam and its two neighbors, admitting sensing only when a candidate exceeds the serving beam by a fixed margin.Source: §I · pp. 1–2
  • In three of four scenarios, feedback-only tracking comes within 0.25 dB of the recorded-sweep oracle, while always-on sensing is harmful; in the remaining scenario, accurate sensing reduces loss when feedback is stressed.Weak or late sensing can be harmful, and probe verification does not fully prevent it.Source: §I · p. 2
  • The paper separates system gain over feedback-only references from sensing’s marginal value relative to the same controller with its sensor disabled.It also accounts for probe savings separately.Source: §I · p. 2

II. RELATED WORK

Prior beam-prediction studies commonly invoke sensing unconditionally and compare against beam training rather than explicit feedback-only tracking. This paper positions its contribution as measuring sensing’s incremental value and admitting it through probe verification with backoff.

A. Always-On Sensing-Aided Beam Prediction

  • The paper’s novelty is the feedback-referenced evaluation framework and incremental sensing-value measurement, rather than a new sensing model.It also presents a controller that admits sensing through probe verification and backs off after a failed trial.Source: A. Always-On Sensing-Aided Beam… · p. 2
  • Prior prediction studies typically invoke the sensor unconditionally and do not report sensing’s incremental value over an explicit feedback-only tracker on measured data with explicit probe accounting.Table I positions prior work by when sensing is invoked, its decision signal, and its baseline.Source: A. Always-On Sensing-Aided Beam… · p. 2, Table I · p. 3
  • The related-work comparison includes sensing-aided prediction and ISAC predictive beamforming as always-on approaches, while budget-scheduled sensing selects acquisition slots using budget and data age.The table marks whether each approach uses on-demand sensing and whether it measures system gain or marginal value.Source: A. Always-On Sensing-Aided Beam… · p. 3

B. Sensor Scheduling for Beam Management

Prior sensing schedulers generally use budget and data age without serving-link measurements to decide whether sensor data are used. Hybrid RF-checking methods exist, but this paper distinguishes its live-tracker comparison and failed-trial backoff.

B. Sensor Scheduling for Beam Management

  • Budget-based sensing schedulers choose acquisition slots using a virtual budget queue and data age, then reuse stale sensing input in non-sensing slots.Their state lacks serving-link residuals, probe outcomes, and tracker state, and they report sensing-rate–accuracy tradeoffs rather than incremental value against feedback-only tracking.Source: B. Sensor Scheduling for Beam Management · p. 3
  • Other update timers and acquisition policies use position, speed, analytic error-growth priors, or belief uncertainty rather than an external sensing arm checked against a live serving link.The comparison also notes that a dual-timescale belief-state policy uses synthetic channels and has no external sensing arm.Source: B. Sensor Scheduling for Beam Management · p. 3

C. Feedback Tracking and RF-Checked Sensing

  • RF checking and adaptive probing have precedent, but the described systems keep sensing always on and accept candidates against a fixed absolute threshold when they check them.This work instead checks whether same-epoch probes beat the live tracker’s serving beam, suspends sensing after a failed check, and carries probe outcomes into future tracker state.Source: C. Feedback Tracking and RF-Checked… · p. 3

III. SYSTEM MODEL AND CAUSAL EVALUATION

The system model treats beam tracking as a sequential decision problem under restricted observations and explicit probe and sensing costs. Evaluation distinguishes controller system gain from sensing’s marginal value against the same controller with its sensor disabled.

A. Measured Link and Beam Model

  • The measured-link model uses recorded uplink powers for a fixed 64-beam receive codebook, and interprets selected measured receive beams as serving beams.Reading the same indices as downlink transmit beams would require calibrated reciprocity and matched codebooks, neither validated by the dataset.Source: A. Measured Link and Beam Model · p. 3

B. The Decision Loop

  • After link setup, each epoch allows candidate probes, a separately costed sensing-model invocation, and a serving-beam commitment; the policy never observes the full power vector.It may read at most one uncharged incumbent or serving-beam power per epoch, while candidate evaluations are charged.Source: B. The Decision Loop · p. 4

C. Objective and the Value of Information

  • Performance is mean power loss against the recorded-sweep oracle, reported alongside additional probes and model invocations per epoch.The loss averages -10 log10 of the capture ratio, rather than applying the logarithm to the average capture ratio.Source: C. Objective and the Value of… · p. 4
  • System gain compares controller loss with an explicitly specified feedback-only reference, while marginal sensing value compares the controller with its sensor disabled and all other components fixed.A positive system gain is credited only when the controller uses no more candidate-beam probes than its reference.Source: C. Objective and the Value of… · p. 4

D. Replay Protocol

The replay limits policies to acquired link observations and uses the full sweep only for scoring. It shows how oracle initialization at every window boundary can make a tracker appear far better than its causal rollout.

D. Replay Protocol

  • Policies read at most one uncharged serving-beam power and only explicitly probed candidate powers; the full 64-beam sweep scores loss and defines the oracle.Link setup is one full sweep whose cost is excluded.Source: D. Replay Protocol · p. 4
  • The replay assumes an exogenous trajectory and same-epoch measurements, so it does not model signaling delays or action-dependent network dynamics.A probed candidate can be selected and applied within the same epoch; one-epoch delays and stale observations are tested separately.Source: D. Replay Protocol · p. 4

E. The Oracle-Feedback Flaw

  • Oracle initialization makes Scenario 32 persistence appear to lose 0.17 dB, versus 2.49 dB in causal rollout—a 2.32 dB difference.Scenarios 33, 31, and 9 show the same pattern: 0.25 versus 3.07, 0.30 versus 1.63, and 4.11 versus 6.35 dB.Source: E. The Oracle-Feedback Flaw · pp. 4–5
  • Because each policy’s past decisions determine its next observations, causal replay makes tracking errors compound as they would in deployment.The oracle-initialized windowed reference silently grants reacquisition at every window boundary.Source: E. The Oracle-Feedback Flaw · pp. 4–5

F. Scenarios, Splits, and Statistical Units

The evaluation uses pass-level test splits and paired bootstrap intervals, reports oracle-relative capture loss, and establishes a feedback-only tracker whose uncertainty drives probing.

F. Scenarios, Splits, and Statistical Units

  • Passes are the statistical unit because frames within a pass are strongly dependent; paired pass-level bootstrap intervals measure differences between fixed policies.The intervals cover pass-to-pass variation, not variation in the design choices behind the policies.Source: F. Scenarios, Splits, and Statistical… · p. 5

G. Metrics and the Decibel Scale

  • Capture-loss exceedance is the fraction of frames whose oracle-relative loss exceeds a threshold, not an outage probability.Top-1 accuracy is not a headline metric because one-bin errors can be nearly free in power while distant errors are catastrophic.Source: G. Metrics and the Decibel Scale · p. 5

A. Innovation-Driven Tracker in Beam-Index Space

  • The feedback-only tracker predicts beam-index drift and grows positional uncertainty from its measured innovation rate, probing when uncertainty exceeds slack or an alarm or refresh occurs.It updates its drift and innovation estimates from its own probe outcomes.Source: A. Innovation-Driven Tracker in… · p. 5
  • The tracker loses 1.045 dB at 3.64 probes/epoch on Scenario 9, with slack selected on validation; Scenario 33 inherits the Scenario-32 setting.The mechanism is not claimed as novel, and any tracker producing serving-beam, probe-outcome, and innovation signals can be substituted into the controller.Source: A. Innovation-Driven Tracker in… · pp. 5–6

B. Baseline Performance Across Regimes

Feedback-only tracking performs very differently across scenarios: it is close to the oracle in three low-innovation cases but leaves substantial loss in Scenario 9. The tracker’s innovation statistic separates these regimes, while the sensing model’s cost and performance vary by site.

B. Baseline Performance Across Regimes

  • 1.045 dB at 3.64 probes/epoch is the Scenario-9 tracker result; the ±3 fixed-cadence baseline reaches 0.80 dB at 6.76 probes/epoch.The ±3 cadence captures at least 90% of available power on 79.0% of frames, versus 69.5% for the tracker.Source: B. Baseline Performance Across Regimes · p. 6
  • At 3.6 probes/epoch, the better posterior-carrying tracker spends 3.00 and loses 5.85 dB, while the Bayesian filter reaches 0.16 dB at 10.4 probes/epoch.Below about three probes/epoch the innovation tracker is the strongest feedback-only policy found; above that, a tracker with a scheduled re-sweep is stronger.Source: B. Baseline Performance Across Regimes · p. 6
  • The other three scenarios’ tracker losses are 0.06–0.25 dB from the oracle at 1.68–2.27 probes/epoch, leaving little mean loss for sensing to reduce.The same-road day/night pair falls in this low-innovation regime; the three daytime scenarios span both regimes.Source: B. Baseline Performance Across Regimes · p. 6

C. Regime Separation in the Innovation Statistic

  • Median innovation is 0.91 in Scenario 9 versus 0.50, 0.36, and 0.37 in the low-innovation scenarios, although upper tails overlap and still cross the trial threshold.The optimal beam’s median net index change is 12.4 indices/s in Scenario 9 and 1.6–3.8 indices/s at the other sites.Source: C. Regime Separation in the Innovation… · p. 6

D. Invocation Cost and What an Invocation Delivers

  • On Scenario 9, the vision arm loses 0.396 dB on covered frames and the radar alternative 0.721 dB; at low-innovation sites the sensing arm loses 0.47–2.49 dB.The low-innovation tracker leaves only 0.06–0.25 dB mean loss to recover.Source: D. Invocation Cost and What an… · p. 6
  • The model is idle on 93–99% of epochs in low-innovation scenarios and 67% in the high-innovation scenario; capture energy and latency are not measured.Reported inference times exclude capture, preprocessing, radio measurement, reporting, and beam application.Source: D. Invocation Cost and What an… · p. 7

E. Degradation at the Other Sites

The controller combines a post-setup re-sweep with probe-verified sensing admission and blacklist backoff. Results show substantial sensitivity to sensing performance across sites, while the controller’s test-informed design chronology and evaluation of confidence-based admission qualify its scope.

E. Degradation at the Other Sites

  • At the other sites, vision top-1 falls to 27.9%, 7.0%, and 2.8% on Scenarios 31, 32, and 33; scene content and training-set size remain confounded.The same-site day-to-night replay increases covered-frame loss from 1.49 to 2.49 dB, while traffic and trajectories also differ.Source: E. Degradation at the Other Sites · p. 7

F. Model-Side Confidence

  • The vision model’s softmax maximum and top-two margin have AUC of only 0.45–0.65 for separating large errors, motivating link-side admission tests.The controller uses tracker innovation as its hand-back signal and leaves admission to the probe test.Source: F. Model-Side Confidence · p. 7

V. PROPOSED CONTROLLER: PROBE-VERIFIED SENSING INVOCATION WITH BLACKLIST BACKOFF

  • The controller probes a predicted beam and its two neighbors, admitting sensing only when a candidate beats the served beam by 0.5 dB; a failed trial triggers a 256-epoch blacklist.It re-sweeps three epochs after setup and periodically or after an alarm rechecks accepted sensing.Source: §V · p. 7
  • Holding out Scenario 9 gives a controller gain of +0.73 dB ([0.23, 1.28]) over the plain tracker; holding out Scenario 31 or 32 gives −0.10 dB ([−0.30, 0.07]) and −0.01 dB.The re-sweep entered the design after test analysis of an earlier constant set, after which constants were re-selected on validation.Source: §V · pp. 7–8

VI. EXPERIMENTAL RESULTS

In high-innovation Scenario 9, the controller improves on feedback-only references, while the marginal value of sensing depends on whether the re-sweep is already present. Setup lock loss explains the concentrated gains; low-innovation scenarios show little system gain.

A. System Gain and the Marginal Value of Sensing

  • In three low-innovation scenarios, the controller loses at most 0.02 dB to feedback-only and invokes the model on 1.0–6.8% of frames, while always-on sensing loses more.The controller also probes more than the plain tracker at these sites, up to 22% more on Scenario 32.Source: A. System Gain and the Marginal Value… · pp. 8–9
  • +0.82 dB over the plain tracker and +0.15 dB over a probe-limited re-sweeping reference, with fewer probes than either, in high-innovation Scenario 9.At the controller’s probe budget, the feedback-only frontier is 0.37 dB higher.Source: A. System Gain and the Marginal Value… · pp. 8–9, Table II · p. 9
  • A scheduled post-setup re-sweep removes every observed lock loss, while the lock-loss-recovery trigger fires too rarely to do so.The relevant re-sweeping reference leaves the controller +0.15 dB better with fewer probes.Source: A. System Gain and the Marginal Value… · p. 9
  • With the re-sweep, sensing cuts probing from 4.12 to 3.25 probes/epoch, a 21% saving, while its +0.04 dB marginal loss reduction is not statistically resolved.The one-sided 95% lower bound is −0.05 dB.Source: A. System Gain and the Marginal Value… · p. 9
  • Seven of 28 test passes gain 2.73–3.63 dB over the plain tracker, while the other 21 average −0.02 dB.A sign test does not reject parity (p = 0.57), although a sign-flip permutation test on the mean does (p = 0.009).Source: A. System Gain and the Marginal Value… · p. 8, Figure 4 · p. 10

B. Attribution of the Gain

The attribution results separate sensing proposals, measured verification, and backoff: verification can reject weak or stale proposals, but it does not guarantee sustained sensing quality. The blacklist, rather than the innovation gate, limits repeated model calls.

B. Attribution of the Gain

  • Replacing the sensor proposal with tracker extrapolation raises Scenario-9 loss by 1.93 dB, and using a random nearby proposal raises it by 2.92 dB.A proposal from 20 frames earlier raises loss by 0.17 dB; verification rejects stale proposals often enough to limit that penalty.Source: B. Attribution of the Gain · p. 10
  • The Scenario-9 lock-loss pass shows the re-sweep returning both controller variants to the best beam, then sensing saving 27 probes at 0.03 dB higher loss until a verified collapse suspends it.On the Scenario-31 pass, the trial fails because feedback is already within the admission margin, and the blacklist then holds.Source: Figure 5 · p. 10
  • Removing the blacklist increases model use from 1.0–33.0% to 30.1–65.9% of frames across the four scenarios; Scenario-33 loss rises from 0.186 to 1.492 dB.An eight-epoch blacklist already restores most of that loss; removing the innovation gate changes loss by at most 0.027 dB.Source: B. Attribution of the Gain · p. 10
  • Verification does not make always-on sensing safe: in the other three scenarios, it remains 0.26–2.17 dB behind feedback-only with or without verification.On Scenario 9, verification on every frame makes always-on sensing not significantly worse than the controller.Source: B. Attribution of the Gain · p. 11
  • Without the admission trial, low-innovation-site loss rises from 0.07–0.19 dB to 0.30–1.35 dB; removing verification and the blacklist raises it to 0.45–2.36 dB.In Scenario 9, the two ablations cost 0.085 and 0.135 dB, and both intervals exclude zero.Source: B. Attribution of the Gain · p. 11

C. What Governs the Gain and the Value of Sensing

Longer decision intervals increase sensing’s value in Scenario 9 when the sensor remains accurate, but weak sensing can add loss elsewhere. The decimation study tests interval sensitivity, not vehicle speed or beam dynamics alone.

C. What Governs the Gain and the Value of Sensing

  • +0.82, +1.88, and +3.66 dB are the Scenario-9 gains over feedback-only at d = 1, 2, and 4, respectively.At d = 4, feedback-only loss is 5.98 dB versus 2.32 dB for the controller.Source: C. What Governs the Gain and the Value… · p. 11, Table IV · p. 12
  • At d = 2 and 4, Scenario-9 marginal value of sensing rises to +1.92 and +3.34 dB, while the weaker arms of Scenarios 32 and 33 cost about 0.6 dB at d = 4.Probe verification does not prevent those losses.Source: C. What Governs the Gain and the Value… · p. 12, Figure 6 · p. 12
  • Across twelve scenario-by-decimation cells, gain correlates more with current-slot sensing accuracy (+0.72) and innovation (+0.53) than with drift (+0.27).The pattern is descriptive because the cells reuse passes from four scenarios.Source: C. What Governs the Gain and the Value… · p. 12
  • Decimation changes feedback timing, prediction coverage, and sensor-to-commit latency, so these results test longer decision intervals rather than vehicle speed or beam dynamics alone.The controller and site settings remain fixed across d; the sensor receives consecutive native frames in the optimistic construction and d-spaced frames in the pessimistic one.Source: C. What Governs the Gain and the Value… · pp. 11–12

D. Robustness of the Operating Point

Stress tests show sensing’s marginal value grows when link feedback is noisy, but stale sensing can worsen loss when link observations are current. Under delayed link information, sensing remains beneficial in the tested settings.

D. Robustness of the Operating Point

  • +2.42 dB is the controller’s gain over the plain tracker at σm = 2 dB, while the marginal value of sensing reaches +1.80 dB over the sensor-disabled controller.Feedback-only degrades faster than the controller under emulated measurement noise.Source: D. Robustness of the Operating Point · p. 12
  • With current link observations, a one-epoch-late sensor raises loss to 0.511 dB and leaves it 0.25 dB above the sensor-disabled controller; two-epoch lateness raises the excess to 0.42 dB.When link observations are one epoch stale, sensing adds +0.57 dB over the sensor-disabled controller, and still adds +0.27 dB when the sensor is also one epoch late.Source: D. Robustness of the Operating Point · pp. 12–13
  • With sensor output and probe outcomes both one epoch late, the controller gains +0.58 dB over feedback-only but falls 0.20 dB behind the sensor-disabled controller.At low-innovation sites, the controller stays within 0.12 dB of feedback-only under every tested delay; end-to-end NR timing compliance is not established.Source: D. Robustness of the Operating Point · p. 13

VII. DISCUSSION AND CONCLUSION

The results make sensing’s value conditional on link-feedback limitations and sensing accuracy and timeliness. The controller improves on feedback-only in the high-innovation scenario, but the evidence is exploratory and bounded by the replay setup and tested scenarios.

VII. DISCUSSION AND CONCLUSION

  • In Scenario 9, the controller gains +0.82 dB over the plain feedback-only tracker and +0.15 dB over the probe-limited re-sweeping reference, with fewer probes than either.Where feedback-only is within 0.25 dB of the oracle, the controller loses at most 0.02 dB to feedback-only while invoking the model on 1.0–6.8% of frames; always-on sensing is harmful there. At Scenario 9’s native interval, sensing does not resolve a loss reduction but substitutes for 21% of candidate-beam probes.Source: §VII · p. 13
  • One validation-selected constant set works without a regime label, and constants selected without Scenario 9 retain a +0.73 dB gain there.Ablations assign current sensing the proposal role, verification a protective role when sensing is weak, and blacklist backoff a role in limiting model calls; the innovation gate is nearly inert.Source: §VII · p. 13
  • The evidence covers four 60-GHz scenarios from three base-station sites on two streets in one dataset family, at low vehicle speeds.The larger gains at longer decision intervals come from decimating cached native-cadence predictions and indicate, rather than measure, what slower feedback would give; transfer to new sites, sensors, models, codebooks, or bands is not established.Source: §VII · p. 13
  • Causal replay relies on exhaustive recorded sweeps and idealized within-epoch timing, while the re-sweep entered the design after test-set analysis exposed setup lock loss.Sequential measurement timing errors cannot be reconstructed, making Scenario-9 comparisons involving the re-sweep test-informed rather than confirmatory.Source: §VII · p. 13
  • Whether probe or inference costs bind was not measured, and verification does not guarantee nonnegative marginal sensing value when operating conditions change.The authors recommend evaluating against feedback-only references with standard refinements and against the same policy with its sensor disabled, reporting probe and invocation costs separately.Source: §VII · p. 13

Related papers

Loading 2610.20196v2…