Focus-Verified Object Discovery for Near-Field Radio Sensing

Very large antenna arrays at millimeter-wave and terahertz frequencies are close enough to the objects they sense that incoming waves arrive curved rather than flat. This regime is called the radiating near field. In it, an array can do more than point a beam in a direction. It can focus energy on a particular point at a chosen distance, the way a lens focuses at a chosen depth. This is beamfocusing. It means a single array can tell apart two objects that sit in the same direction at different ranges. Today, though, the learning-based work that turns these curved-wave signals into object detections is fully supervised and studies one object at a time. Learning to find many objects without labels has been developed for cameras and LiDAR instead. Those methods rely on motion: things that move together belong together. They are blind to objects that never move.

We propose to train a class-agnostic, multi-object detector for near-field radio channels without any labels, using two complementary teachers. The first is a motion teacher. It finds moving scatterers from slow-time phase changes and uses them as initial pseudo-labels, following the moving-to-mobile recipe from LiDAR. The second is the ingredient no camera or LiDAR pipeline has: an active focus probe. When the detector proposes a candidate location that never moved, the array refocuses on that exact range-angle point. The candidate is accepted as a pseudo-label only if the focused return is compact in range as well as angle. The array can change its own illumination, so it can test static hypotheses physically instead of relying on appearance generalization alone. In plain terms, motion tells the system what objects look like, and the ability to focus at depth lets it check its guesses about things that sit still.

Core Claim

Near-field beamfocusing turns object discovery from a passive problem into an active one. Our claim is that a student detector trained on motion seeds, and refined by focus-verified static pseudo-labels, will do two things that motion-only self-training transferred to radio cannot. It will recover never-moving objects, such as debris or permanently parked vehicles. It will also keep co-directional targets separated as the number of objects grows. Both gains should vanish when the same array is restricted to far-field (flat-wave) beams. That makes the curvature of the wavefront the testable cause of the improvement.

The focus-consistency score we use is simple. Let $\mathbf{h}$ be the received channel snapshot across the $N$ array elements. Let $\mathbf{a}(r,\theta)$ be the near-field steering vector, meaning the per-element phases that match a spherical wave from range $r$ and angle $\theta$. A candidate at $(r_0,\theta_0)$ is scored by

$$ s(r_0,\theta_0) = \frac{|\mathbf{a}(r_0,\theta_0)^{H}\mathbf{h}|^2}{\max_{|r-r_0|>\delta}\,|\mathbf{a}(r,\theta_0)^{H}\mathbf{h}|^2}, $$

where $\delta$ is a few range-resolution cells. The score is high when energy concentrates at that depth along that direction. It is low when energy is smeared along the ray, as happens with extended walls or diffuse multipath. Far-field beams have no range selectivity, so their analogue of $s$ is identically uninformative. This is what the far-field ablation exploits.

Positioning Against Prior Work

The near-field side supplies the physics and the strongest supervised models, and it motivates the gap. Beam Focusing for Near-Field Multi-User Localization shows that wavefront curvature lets one array estimate both angle and range for several users. Its two-stage method alternates location estimates with refocusing weights, which is the closest existing use of "refocus to verify." However, the users there are active transmitters sending pilots, and their number is known in advance. We address passive scatterers of unknown count with no cooperation. A Spatio-Temporal-Frequency Transformer Framework for Near-Field Target Recognition and Extended-Target Classification and Localization for Near-Field ISAC show that near-field channel tensors carry rich shape and position signatures of extended vehicles. They do so from full-wave simulations of ShapeNet models, but each scene contains one isolated target and all training is supervised. We reuse their channel-tensor encoders and their simulation methodology, and we drop both the labels and the single-target assumption. The wideband beamfocusing work in Beamfocusing Optimization for Near-Field Wideband Multi-User Communications constrains our design. Focal spots drift across subcarriers, so the probe must use per-subcarrier steering vectors, not one shared focus.

The discovery side supplies the training recipe and its known failure. LISO clusters residual scene flow into boxes, smooths their trajectories, and self-trains a single-frame detector that then finds parked cars. MOD-UV does the same from monocular video and adds synthetic shrinking of near objects to help with distant ones. Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D motion passes pseudo-labels back and forth between camera and LiDAR. All three generalize from moving to stationary instances of categories that moved at some point. All three remain blind to categories that never moved, as Unsupervised Learning from Video to Detect Foreground Objects in Single Images and Unsupervised Object Learning via Common Fate describe. Our focus probe is a second, motion-independent source of supervision that targets exactly this blindspot. It differs from appearance-only generalization because it acts on the scene. Video Class Agnostic Segmentation Benchmark for Autonomous Driving motivates why never-moving novel obstacles matter.

Research Program

No multi-target near-field sensing benchmark exists in the literature we found, so the first deliverable is one. We will place a simulated extremely large array at the roadside. It will be driven by real scene layouts and trajectories from Argoverse 2 and the Waymo Open Dataset. Each annotated object is replaced by a ShapeNet mesh of matching size. We then add never-moving inanimate objects (debris, cones, crates) that no motion teacher can seed.

Scattering is computed at two fidelities. The first is a fast physical-optics or point-scatterer approximation, used for training at scale. The second is full-wave volume-integral-equation simulation, as in the STF and dual-branch papers, on a smaller set of evaluation scenes. Comparing the two quantifies how much our conclusions depend on the cheap simulator. We vary array aperture, the number of sensing subcarriers $K$ (from 1 to 16, matching the sparse-subcarrier regime of prior ISAC work), multipath richness, and object density.

The student network maps one channel snapshot to a range-angle heatmap of object centers with extents. Its encoder follows the STF Transformer's spatial-convolution-plus-attention design. Training then proceeds in three steps.

In the first step, the motion teacher removes the static channel component by slow-time differencing. It clusters the residual energy in range-angle space and links clusters over time with smooth-trajectory regularization, as LISO does. These tracks become the initial pseudo-labels.

In the second step, the student is trained on single snapshots. It therefore has to learn what moving objects look like in the curved-wave signal, and it starts firing on stationary objects of seen categories.

In the third step, the student's confident detections that have no motion support are sent to the focus probe. The array is retuned to $(r_0,\theta_0)$ under the same hardware constraint as the experiment (fully digital, or a 10-RF-chain dynamic metasurface as in the multi-user localization paper). The score $s$ is computed, and candidates above threshold join the pseudo-label set. Because the probe also scores a coarse grid where the student is silent, it can seed never-moving categories directly, not only confirm the student's guesses. The loop repeats.

The probe costs illumination time, so we also study a budgeted version. In it, the probe chooses which candidates to test by student uncertainty. This connects the work to resource limits in multi-target scheduling.

Baselines

The naive reference point is classical radar processing. It uses static-clutter subtraction followed by constant-false-alarm-rate detection on the near-field range-angle map, with a fixed threshold, and involves no learning. By construction it only reports moving objects. It shows what any learned method must add, and it is the motion teacher used without a student. It runs on identical channels. To beat it, the learned methods must recover still objects at all and must lower missed detections in co-directional pairs.

The strongest existing self-supervised method is LISO's three-stage recipe transferred to near-field channels: motion-derived pseudo-labels, trajectory regularization, and iterative self-training of a single-frame detector. We give it the same student architecture, the same unlabeled data, the same number of self-training rounds, and the same hyperparameter search budget. The only difference is that it has no focus probe. It is the strongest competitor because it already solves "moving to mobile," and the literature number to keep in mind is its small moving-versus-still gap on Waymo LiDAR, 0.350 versus 0.322 AP (LISO). That figure comes from a different sensor and is cited only to show what a good moving-to-still transfer looks like, not as a directly comparable target. To beat it, our method must match LISO-RF on seen categories and clearly exceed it on never-moving categories, where we expect LISO-RF recall near zero.

The supervised upper reference is the dual-branch extended-target network of Extended-Target Classification and Localization for Near-Field ISAC, extended with a multi-object detection head and trained on simulator labels. On single isolated targets at $K=4$ that paper reports a planar MAE of 1.1826 m and Succ@1m of 59.35%. We will first reproduce that single-target operating point on our simulator to check calibration, then train on multi-object scenes. For the multi-user localization approach, no reported result was found on Lacuna for localization error as a number; the file only states that a 10-RF-chain dynamic metasurface matches a 500-RF-chain digital array. We therefore run its alternating-projection method ourselves, giving it the true object count as it requires, and also with a count estimated by thresholding.

Two ablations isolate the key ingredient. The first is focus probe only, with no motion seeding. The second, which matters most, is the full method with every steering vector replaced by its far-field (flat-wave) counterpart on the same array, so the probe has no range selectivity. If the far-field ablation performs as well as the near-field version, our claim is false.

Why not just…?

Why not just train supervised on simulator labels, since labels are free in simulation? In simulation this is a real threat, and the supervised dual-branch baseline will likely win on held-in conditions. The argument for self-supervision is deployment. Real installations have no labels, and the multipath and clutter limitations noted for near-field models mean the simulator's curvature will not match reality. Our test trains the supervised model on the cheap simulator and evaluates on full-wave scenes with added multipath and object categories it never saw. The self-supervised methods instead train on unlabeled data from the target condition. If the supervised model still wins under that shift, we will report self-supervision as useful only for adaptation, not as a replacement.

Why not just get range from bandwidth, as ordinary radar does, and skip near-field focusing? This is also a genuine competitor. With wide bandwidth, plain range bins already separate co-directional objects, and the curvature adds little. Our position is narrower. Integrated sensing-and-communication systems sense on a handful of subcarriers, which is the $K=1$ to $K=8$ regime the STF and dual-branch papers study. With so few subcarriers, bandwidth gives poor range resolution while the aperture's curvature still gives fine depth selectivity. We will sweep $K$ against aperture and report the crossover where the far-field-plus-bandwidth ablation catches up. That defines exactly where the method is worth using.

Why not just rely on appearance generalization, as LISO and MOD-UV do for parked cars? For stationary instances of categories that moved, this will probably work, and LISO-RF may match us there. It cannot work for categories with zero motion examples. A student trained only on motion seeds has never been rewarded for firing on them, which is the blindspot formalized by Croitoru et al. The never-moving debris split of our benchmark is built to measure exactly this difference. If LISO-RF recovers those objects anyway, through some accidental similarity, the probe's contribution reduces to cleaning up co-directional pairs, and we will say so.

Proposed Experiments and Decisive Checks

We evaluate class-agnostic detection in the range-angle plane. The metrics are average precision with a center-distance match, and success within 1 m, defined as in the dual-branch paper for comparability. Both are reported separately for moving objects, stationary objects of moving categories, and never-moving categories, and binned by range and object density. Multi-target scalability is measured three ways: identity switches across tracked co-directional pairs, recall as objects per scene grow from 1 to 20, and error in the estimated object count. For context with optical and LiDAR discovery, we note the reported xMOD recall of about 4.1% beyond 30 m and its KITTI late-fusion F1 of 27.4 (Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D motion). These come from different sensors and are not head-to-head targets.

Three outcomes would decide the project. First, on never-moving categories the full method must beat LISO-RF by a wide margin, and the probe-only ablation should recover a substantial part of that gain. Second, the far-field steering ablation must lose that gain, and must lose separation of co-directional pairs, in the few-subcarrier regime; the gap should shrink as $K$ grows. Third, the ranking of methods on the cheap simulator must hold on the full-wave evaluation scenes; otherwise the result is a simulator artifact. Finally, we will report how many probe dwells per scene the budgeted variant needs to reach most of the unbudgeted recall. This tells us whether active verification is affordable alongside communication traffic.