Sensing-Assisted Communication

Using Scene Knowledge for Communication

So far in this chapter we have seen that channel estimation is imaging (Section 30.1) and that distributed nodes can cooperatively image the environment (Section 30.2). The natural next question is: what can you DO with the reconstructed scene?

The answer is sensing-assisted communication (SAC): using the reconstructed environment map β€” the digital twin β€” to predict channels, select beams, plan handovers, and allocate resources WITHOUT the overhead of exhaustive pilot-based estimation. This closes the loop: imaging feeds communication, which in turn provides the waveforms for further imaging.

Definition:

Sensing-Assisted Communication

In SAC, the communication system exploits sensing-derived scene knowledge to reduce channel estimation overhead:

H^(t)=F ⁣(St; xbs, x^ue(t), f)\hat{\mathbf{H}}(t) = \mathcal{F}\!\left(\mathcal{S}_t;\, \mathbf{x}_{\mathrm{bs}},\, \hat{\mathbf{x}}_{\mathrm{ue}}(t),\, f\right)

where St\mathcal{S}_t is the scene state (digital twin), x^ue(t)\hat{\mathbf{x}}_{\mathrm{ue}}(t) is the UE position estimated from sensing, and F\mathcal{F} is the forward renderer.

The predicted channel H^\hat{\mathbf{H}} enables:

  • Beam selection: vβˆ—=arg⁑max⁑v∣H^Hv∣2\mathbf{v}^* = \arg\max_{\mathbf{v}} |\hat{\mathbf{H}}^H \mathbf{v}|^2 without beam sweeping.

  • Pilot reduction: skip pilots when βˆ₯H^βˆ’Hprevβˆ₯F<Ο΅\|\hat{\mathbf{H}} - \mathbf{H}_{\mathrm{prev}}\|_F < \epsilon (scene stable).

  • Proactive handover: predict channel quality at neighbouring cells from the scene model before the UE reports measurements.

  • Blockage prediction: detect approaching blockers (vehicles, pedestrians) from the sensing returns and pre-compute alternative beam paths.

The digital twin acts as a channel predictor: given the scene geometry and the UE position, it renders the expected channel without any pilot overhead. Pilots are used only to refine the prediction (residual estimation), reducing overhead dramatically.

,

Definition:

RF Digital Twin for Communication

An RF digital twin T=(S,F,U)\mathcal{T} = (\mathcal{S}, \mathcal{F}, \mathcal{U}) consists of:

  1. Scene state St\mathcal{S}_t: a representation (neural or geometric) encoding geometry, material properties, and dynamic objects.

  2. Forward renderer F\mathcal{F}: maps scene state to predicted RF observables: y^=F(St;s,r,f)\hat{\mathbf{y}} = \mathcal{F}(\mathcal{S}_t; \mathbf{s}, \mathbf{r}, f).

  3. Update mechanism U\mathcal{U}: ingests new measurements and updates the scene: St+1=U(St,yt+1)\mathcal{S}_{t+1} = \mathcal{U}(\mathcal{S}_t, \mathbf{y}_{t+1}).

The twin operates in a continuous sense-update-predict loop with latency budget less than the channel coherence time TcT_c.

,

Theorem: Pilot Overhead Reduction via Digital Twin

For a MIMO-OFDM system with NtN_t antennas, KK subcarriers, and coherence time TcT_c slots, the pilot overhead fraction is:

Ξ·pilot=NtTc\eta_{\mathrm{pilot}} = \frac{N_t}{T_c}

with full channel estimation. With the digital twin predicting the channel and pilots used only for residual estimation:

Ξ·DT=NtTcβ‹…rank(Hβˆ’H^DT)rank(H).\eta_{\mathrm{DT}} = \frac{N_t}{T_c} \cdot \frac{\mathrm{rank}(\mathbf{H} - \hat{\mathbf{H}}_{\mathrm{DT}})}{\mathrm{rank}(\mathbf{H})}.

If the twin captures the dominant LL out of PP paths, the overhead reduction factor is ∼L/P\sim L/P.

The channel lies in a subspace of dimension PP (number of paths). The digital twin predicts LL of these, leaving a residual of rank Pβˆ’LP - L. Estimating the residual requires only Pβˆ’LP - L pilot dimensions instead of NtN_t.

Proof
Subspace decomposition

Decompose H=H^DT+Ξ”H\mathbf{H} = \hat{\mathbf{H}}_{\mathrm{DT}} + \Delta\mathbf{H}, where H^DT\hat{\mathbf{H}}_{\mathrm{DT}} captures LL paths (rank=L\mathrm{rank} = L) and Ξ”H\Delta\mathbf{H} has rank≀Pβˆ’L\mathrm{rank} \leq P - L.

Pilot requirement for residual

Estimating Ξ”H\Delta\mathbf{H} of rank Pβˆ’LP - L requires M=Pβˆ’LM = P - L pilot dimensions (compressed sensing or subspace tracking). The overhead fraction becomes Ξ·DT=(Pβˆ’L)/Tc=Ξ·pilotβ‹…(Pβˆ’L)/P\eta_{\mathrm{DT}} = (P-L)/T_c = \eta_{\mathrm{pilot}} \cdot (P-L)/P. Since rank(H)≀P\mathrm{rank}(\mathbf{H}) \leq P and the twin captures LL paths: Ξ·DT/Ξ·pilot=(Pβˆ’L)/P\eta_{\mathrm{DT}}/\eta_{\mathrm{pilot}} = (P-L)/P. β– \blacksquare

,

Beam Prediction Accuracy from Reconstructed Map

Compare beam selection accuracy: exhaustive beam sweeping (all beams measured) vs. digital twin prediction (top-kk candidates measured).

The plot shows the probability of selecting the optimal beam and the effective throughput as a function of DT scene accuracy (NMSE of the reconstructed map).

Observe that even a moderate DT accuracy (βˆ’10-10 dB NMSE) achieves >90%> 90\% correct beam selection with only k=3k = 3 measurements instead of NbeamsN_{\mathrm{beams}}.

Parameters
64
3
100

Example: Digital Twin-Aided Beam Management

A 64-antenna 28 GHz base station uses a codebook of 64 beams. Standard beam sweeping requires 64 pilot slots. A digital twin predicts the top-3 beam candidates. Compare overhead and throughput with coherence time Tc=100T_c = 100 slots.

Solution
Standard beam sweeping

All 64 beams measured: 64 pilot slots. Optimal beam guaranteed. Overhead fraction: 64/100=64%64/100 = 64\%. Data slots: 36.

DT-aided (top-3)

Only 3 beams measured: 3 pilot slots. If the DT top-3 is correct (probability ∼92%\sim 92\% with scene NMSE <βˆ’10< -10 dB): optimal beam. If not: fall back to full sweep (detected via low SINR). Effective overhead: 0.92Γ—3+0.08Γ—64=7.90.92 \times 3 + 0.08 \times 64 = 7.9 slots.

Throughput gain

Standard: 3636 data slots. DT-aided: 100βˆ’7.9=92.1100 - 7.9 = 92.1 data slots. Throughput increase: 92.1/36=2.56Γ—92.1/36 = 2.56\times.

Definition:

The Sensing-Reconstruction-Prediction Loop

The full sensing-assisted communication pipeline forms a closed loop:

  1. Sense: ISAC waveforms illuminate the environment and serve communication users simultaneously.

  2. Reconstruct: The sensing returns are processed by the imaging pipeline (LASSO, OAMP, or learned methods from Parts IV--VI) to update the digital twin St\mathcal{S}_t.

  3. Predict: The twin renders predicted channels H^(t+1)\hat{\mathbf{H}}(t+1) for all active users.

  4. Communicate: Beamforming vectors and resource allocation are computed from the predicted channels, closing the loop.

The loop latency must satisfy Ξ”tloop<Tc\Delta t_{\mathrm{loop}} < T_c (channel coherence time) for the predictions to be actionable.

,

Common Mistake: Scene Staleness in Sensing-Assisted Communication

Mistake:

Using a digital twin that was last updated several coherence times ago to predict the current channel.

Correction:

The twin lags behind reality by the update latency. Fast-moving objects (vehicles at 100 km/h travel 2.8 m per 100 ms update) cause the predicted channel to diverge rapidly.

Mitigation:

  1. Motion prediction: extrapolate object positions using tracked velocities from the Kalman filter.
  2. Adaptive pilot insertion: when the predicted channel deviates from the measured channel by more than a threshold, trigger a full pilot burst.
  3. Freshness-aware scheduling: prioritise pilot resources for UEs in rapidly changing environments.

Common Mistake: Position Accuracy for Coherent Beam Prediction

Mistake:

Assuming the digital twin can predict the channel with sub-wavelength accuracy, enabling coherent beamforming without any pilots.

Correction:

Coherent beamforming requires phase-accurate channel prediction, which demands UE localisation within Ξ»/2\lambda/2. At 28 GHz, this is ∼5\sim 5 mm β€” far beyond current positioning accuracy (metres for GPS, sub-metre for 5G positioning).

In practice, the twin predicts dominant path directions (angular domain) rather than exact phases. This suffices for beam selection but not for coherent precoding. Residual pilots are still needed for phase refinement.

Quick Check

A digital twin captures 4 out of 5 propagation paths. By approximately what factor is the pilot overhead reduced?

5Γ—5\times (from Ξ·\eta to Ξ·/5\eta/5)

4Γ—4\times (from Ξ·\eta to Ξ·/4\eta/4)

1.25Γ—1.25\times (from Ξ·\eta to 4Ξ·/54\eta/5)

Correction:
5Γ—5\times (from Ξ·\eta to Ξ·/5\eta/5)

The residual has rank 1 (one uncaptured path) out of 5 total. The overhead reduces by a factor of P/(Pβˆ’L)=5/1=5P/(P-L) = 5/1 = 5.

Why This Matters: Beam Management in 5G NR and Beyond

In 5G NR, beam management (SSB-based beam sweeping, CSI-RS-based refinement) consumes significant overhead at mmWave frequencies β€” up to 60% of the coherence interval for large codebooks. The sensing-assisted approach of this section can reduce this to <10%< 10\% by narrowing the beam search to a small set of DT-predicted candidates. This is actively being studied for Release 19/20 of 3GPP, where "environment-aware" beam management is a study item.

⚠️Engineering Note

Latency Budget for Sensing-Assisted Communication

The sensing-reconstruction-prediction loop must complete within the channel coherence time TcT_c. At 28 GHz with pedestrian mobility (3 km/h): Tcβ‰ˆ50T_c \approx 50 ms. With vehicular mobility (60 km/h): Tcβ‰ˆ2.5T_c \approx 2.5 ms.

Budget allocation:

  • Sensing: 1 OFDM symbol β‰ˆ70\approx 70 ΞΌ\mus.
  • Reconstruction (OAMP, 10 iterations): β‰ˆ5\approx 5 ms for a 128Γ—128128 \times 128 image on GPU.
  • Channel prediction (forward render): β‰ˆ1\approx 1 ms per UE.
  • Beam computation: β‰ˆ0.5\approx 0.5 ms.

Total: β‰ˆ7\approx 7 ms β€” feasible for pedestrian scenarios (Tc=50T_c = 50 ms) but tight for vehicular (Tc=2.5T_c = 2.5 ms). For high-mobility, simplified reconstruction (backprojection + Kalman tracking) is needed.

Practical Constraints
  • β€’

    Pedestrian coherence time: ∼50\sim 50 ms (28 GHz)

  • β€’

    Vehicular coherence time: ∼2.5\sim 2.5 ms (28 GHz)

  • β€’

    GPU inference latency for OAMP: ∼5\sim 5 ms

RF Digital Twin

A real-time virtual replica of the physical RF environment, consisting of a scene representation, a forward renderer, and an update mechanism. The twin predicts RF observables (channels, sensing returns) from the scene geometry, enabling communication optimisation without exhaustive measurement.

Related: {{Ref:Def Digital Twin Comm}}

Sensing-Assisted Communication

A communication paradigm where sensing-derived environment knowledge (e.g., from a digital twin) is used to reduce pilot overhead, predict beams, and optimise resource allocation. The sensing returns provide side information that reduces the effective channel dimensionality.

Related: {{Ref:Def Sensing Assisted Comm}}

Key Takeaway

Sensing-assisted communication closes the loop between imaging and communication: the reconstructed scene (digital twin) predicts channels without pilots, reducing overhead by ∼90%\sim 90\% in typical scenarios. The key is the sensing-reconstruction-prediction loop, which must complete within the coherence time. Beam management is the most immediate application, with 22--3Γ—3\times throughput gain from DT-aided top-kk beam prediction.

Wavefield Networked SensingJoint Communication-Sensing-Imaging Optimization