Learning Target Geometry from Communication Payload Echoes in Near-Field ISAC

A base station that both talks to users and senses its surroundings has to split its frequency band. A few subcarriers (individual frequency tones in an OFDM signal) carry known, deterministic radar probes. The rest carry random user data. Current learned sensing systems read only the dedicated probe tones and treat the data tones as lost to sensing. Yet the data tones bounce off the same car or motorcycle. They carry the same geometry, mixed with fluctuations from symbols that a separate (bistatic) sensing receiver does not know. We propose to treat the two kinds of tones as two "modalities" of one scene and to borrow a tool from multimodal learning: feature disentanglement, which means learning one representation of what both modalities share and a separate one for what is private to each. The shared part should hold target geometry and class. The private part should absorb the payload-induced variation on the data tones and the range ambiguity of having only a few probe tones.

If this works, the sensing-communication trade-off curve shifts. A system could reach the same localization error with fewer dedicated probe tones, so more of the band carries data. The approach also offers a concrete, testable neural reading of the information-theoretic idea that a communication signal can carry state information "for free," with a separate sensing-only refinement on top.

Core Claim

In a bistatic near-field multiple-input multiple-output (MIMO) OFDM link, the receiver sees probe tones $y^{s}_k = H_k p_k + n_k$ with known probes $p_k$. It also sees data tones $y^{c}_k = H_k x_k + n_k$, where the symbols $x_k$ are unknown and random. Here $H_k$ is the near-field channel of the target at tone $k$, and $n_k$ is noise. The geometry in $H_k$ is common to both streams. The randomness of $x_k$ is private to the data stream and, over short bursts, acts as self-interference.

We claim that an encoder trained to separate a shared subspace from private subspaces can recover more localization accuracy from the data tones than either of the two obvious routes. The first route is classical: decode the symbols and divide them out. The second is to compute payload-averaged spatial covariances. The advantage should be largest exactly where Sensing Mutual Information with Random Signals in Gaussian Channels locates the "randomness penalty": few frames and low-to-medium signal-to-noise ratio (SNR). That paper analyzes this penalty for precoder design and shows it vanishing as the number of frames grows. We use its prediction as a falsifiable signature for a learned receiver. The advantage should shrink toward zero as the slow-time burst length grows.

The analogy to theory comes from An Achievable Rate-Distortion Region for Joint State and Message Communication over Multiple Access Channels. That paper splits auxiliary variables into message-bearing variables ($T$) and sensing-only refinement variables ($V$). The latter are decoded after the message and add nothing to the data rate. In our design, the data tones play the role of $T$ while also carrying state, and the few probe tones play the role of $V$. The analogy motivates the design. We do not claim it proves anything.

How This Differs From Existing Work

Two papers on the ShapeNet-derived full-wave near-field dataset are closest to this proposal. Extended-Target Classification and Localization for Near-Field ISAC builds a dual-branch network with a semantic branch for classification and a geometric branch for localization, linked by cross-task attention, and shows good sensing with only $K=4$ sensing tones. A Spatio-Temporal-Frequency Transformer Framework for Near-Field Target Recognition applies a transformer over antenna, slow-time, and sparse-subcarrier tokens.

Both read only the dedicated sensing tones and count communication purely as "subcarriers saved." The first paper's split is by task (what versus where) within one input, not by signal source, and it applies no orthogonality or moment-matching constraint. We keep their outputs and change the input and the decomposition. Every tone is observed, and the split separates scene content shared across tone types from payload-specific content.

The disentanglement machinery comes from multimodal sentiment analysis. MISA projects each modality into a shared and a private subspace, using three losses:

  • central moment discrepancy (matching the first few statistical moments of the shared codes across modalities);
  • an orthogonality penalty between subspaces;
  • a reconstruction loss so that nothing is discarded.

DecAlign adds prototype-level optimal-transport alignment. It also flags that alignment as expensive and that loss weights are hard to tune.

We reuse MISA's three losses, which are cheap, and skip optimal transport for physical-layer latency reasons. That choice is also motivated by the latency concerns raised in Generative AI for Integrated Sensing and Communication: Insights From the Physical Layer Perspective. Multimodal Mixture-of-Experts for ISAC in Low-Altitude Wireless Networks separates sensors into separate experts but never separates shared from private content inside a stream. It is a related design choice, not a competitor on this question.

Research Program

We first build a waveform layer on top of the full-wave channel tensors of the ShapeNet-derived near-field dataset. For each scene, a total band of $K_{tot}$ tones is split into $K$ probe tones and $K_{tot}-K$ data tones carrying random QAM symbols. The data tones are observed bistatically over $N_p$ slow-time frames, with $N_p$ from 1 to 8, matching the STF paper's range, and SNR is swept. Because the channels are precomputed, we can draw many independent payloads for the same geometry at no extra solver cost.

This payload resampling is the key training signal. Two views of one scene with independent payloads must map to the same shared code. The private data-stream code must explain the difference between them.

The model has three parts:

  • a probe-tone encoder and a data-tone encoder, each with the spatial-CNN-plus-transformer backbone of the STF paper;
  • shared and private projection heads on each encoder;
  • the semantic and geometric heads of the dual-branch paper, reading only the shared code plus the probe-private code.

Training adds four terms to the task loss. Central moment discrepancy aligns the probe and data shared codes. A cross-payload consistency term ties the shared codes of resampled views. An orthogonality penalty separates shared from private codes. Reconstruction of both inputs from shared plus private codes keeps the representation complete.

We will also report what each subspace encodes, using linear probes for range, azimuth, class, and the payload symbols themselves. "Disentangled" will then be a measured property, not a label.

Baselines

The naive reference is dedicated tones only. We retrain the dual-branch extended-target network exactly as described, on the $K$ probe tones, ignoring the data tones. At $K=4$ it reports planar mean absolute error (MAE) of 1.1826 m, success within 1 m of 59.35%, and classification accuracy around 99.32% (Extended-Target Classification and Localization for Near-Field ISAC). We will reproduce it in our split before comparing, because our scenes and noise differ. Its sibling is the STF transformer on probe tones. That paper reports 98.6% accuracy at $(N_p, K_{sel})=(8,8)$ and 93.9% at $(1,1)$ (A Spatio-Temporal-Frequency Transformer Framework for Near-Field Target Recognition), and we run it the same way.

The strongest competitors read the data tones without disentanglement. All of them share our backbone, our data, and an equal hyperparameter-search budget measured in GPU-hours.

The monolithic model concatenates all tones into one STF transformer. This tests whether "more tones" alone explains any gain.

The covariance model feeds per-tone sample spatial covariances of the data tones. Their expectation is independent of the payload, so this is the physics-motivated answer to unknown symbols.

The decision-directed model runs a classical linear-MMSE detector on the data tones, divides out the hard-decided symbols, and feeds the resulting channel estimates to the network.

Above all of these sits a monostatic oracle in which the symbols are known and divided out exactly. It bounds how much the data tones can possibly contribute.

There are two ablations. The first is invariance only: keep the cross-payload consistency loss but drop the private heads, orthogonality, and reconstruction. It isolates whether the shared/private structure matters beyond payload augmentation. The second is no moment matching, which removes the probe-data alignment term.

We will count the proposal as beating the field under one condition. At a fixed communication rate, meaning a fixed number of data tones, it must achieve lower planar MAE and higher success within 1 m than the best of the monolithic, covariance, and decision-directed baselines. This must hold with non-overlapping 95% confidence intervals over five seeds, in the short-burst regime ($N_p \le 2$, low-to-medium SNR).

Why not just…?

Why not just feed every tone into one big transformer? This is the cheapest version of "use the data tones," and it may capture much of the gain. Work across both fields reports that monolithic fusion lets a dominant or cleaner stream overshadow a noisier one, and that early sharing hurt the near-field localization branch (Extended-Target Classification and Localization for Near-Field ISAC, DecAlign). Neither result was obtained with random-payload tones, however. The monolithic baseline settles this directly. If it matches our model, the disentanglement story is wrong and we will say so.

Why not just compute covariances, which average the payload away? This is the real threat. With long bursts it should be nearly optimal, and we expect it to catch up as $N_p$ grows. Our case rests on short bursts, where sample covariances retain payload fluctuation. This is the finite-frame randomness penalty that Sensing Mutual Information with Random Signals in Gaussian Channels shows matters at low-to-medium SNR. If the covariance baseline equals our model even at $N_p=1$, we will drop the learned private branch and recommend covariance features instead.

Why not just decode the data and divide it out? At high SNR this approaches the monostatic oracle, and we expect it to win there. At low SNR, decision errors inject structured noise into the channel estimates. The SNR sweep will show where the two curves cross. A proposal that only helps below that crossing is still useful, but we will report the crossing honestly.

Why not just give sensing more dedicated tones? That is the orthogonal-slicing answer, and it pays in rate. The STF results already show what scarcity costs: accuracy falls from 98.6% to 93.9% when tones and frames drop to one. Our rate-versus-MAE curves make this baseline a point on the same trade-off plot.

Proposed Experiments and Decisive Checks

The main result is a family of trade-off curves: planar MAE and success within 1 m against data-tone count (communication rate), for $K \in \{1,2,4,8\}$, $N_p \in \{1,2,4,8\}$, and an SNR sweep, on the ShapeNet-derived near-field dataset. Classification accuracy is reported alongside. There is no reported result on Lacuna for any learned receiver that senses from random-payload tones, so all data-tone baselines are run by us.

Three checks decide the claim.

First, the gain over the covariance and decision-directed baselines must be largest at small $N_p$ and low-to-medium SNR, and must shrink as $N_p$ grows, as the randomness-penalty theory predicts. A gain that stays flat across $N_p$ would point to an artifact rather than the proposed mechanism.

Second, linear probes must show that payload symbols are decodable from the private data code and near chance from the shared code, while range and azimuth show the reverse.

Third, the invariance-only ablation must fall measurably short of the full model. Otherwise the private subspace is decoration, and the contribution reduces to payload augmentation, which we would then report as the finding.