RIS-Aided Wireless Amodal Sensing

• RIS-aided wireless amodal sensing is a technique that integrates programmable metasurfaces with wireless systems to reconstruct complete 3D representations of occluded objects.
• It employs dynamic RIS phase optimization and compressed sensing combined with generative diffusion models to recover hidden structural details from limited measurements.
• Experimental results show improved spatial resolution and significant error reduction, highlighting its potential in privacy-preserving and remote sensing applications.

Reconfigurable Intelligent Surface (RIS)-aided wireless amodal sensing refers to the integration of programmable electromagnetic metasurfaces—RIS—with wireless sensing systems to reconstruct comprehensive (amodal) representations of partially occluded or hidden objects. This paradigm leverages the ability of RIS to control radio propagation dynamically, enabling the recovery of 3D object shapes from limited, often single-view and occlusion-impaired wireless measurements. Practical applications range from robust scene reconstruction in adversarial environments to privacy-preserving sensing and remote perception where direct observation is obstructed. Core advances relate to the use of RIS for enhancing spatial resolution, generating reflection paths that bypass occlusions, and employing advanced learning-based reconstruction, notably with generative deep models (Wang et al., 2 Feb 2026, Kompostiotis et al., 28 Nov 2025).

1. System Model and Channel Formulation

RIS-aided wireless amodal sensing systems typically comprise:

• Transmitter (Tx): Emits known waveforms, generally from a single-antenna or MIMO source.
• Reconfigurable Intelligent Surface (RIS): A planar array with $M$ or $N$ elements, each able to impose a tunable phase shift, thereby shaping the impinging wavefront toward desired spatial sectors (Kompostiotis et al., 28 Nov 2025).
• Region of Interest (ROI): Discretized into $N = N_x \times N_y \times N_z$ voxels; an unknown subset occupied by object(s), some of which may be occluded (Wang et al., 2 Feb 2026).
• Receiver (Rx): Collects the reflected and backscattered signals, typically with a single antenna.

The backscattered signal at Rx is expressed as

$r = q^T (H_r \odot V)\, \omega_v + z,$

where $q$ is the RIS phase configuration, $H_r$ is the RIS-to-ROI channel matrix, $V$ encodes path occlusions ($V_{m,n} \in \{0,1\}$), $\omega_v$ the vector of visible voxel scattering coefficients, and $z$ additive Gaussian noise [(Wang et al., 2 Feb 2026) §II.C].

RIS phase shifts are represented as discrete values $\phi_m \in \mathcal{F}$, where $\mathcal{F}$ is the $2^b$-point quantization grid [$\mathcal{F}$: phase codebook, $b$: bit resolution]. The end-to-end channel model for RIS-enabled sensing explicitly combines free-space gains, RIS-imposed phase, and occlusion.

In MIMO or extended geometries, the composite link is

$y = h_d x + h_r^T \Theta G x + n,$

with $G$ the Tx$\to$RIS channel, $h_r$ the RIS$\to$Rx channel, $\Theta$ the RIS phase matrix, and $n$ noise [(Kompostiotis et al., 28 Nov 2025) §1.2]. The RIS thus creates controllable indirect paths circumventing occlusions.

2. Amodal 3D Reconstruction Pipeline

The process of RIS-aided amodal sensing for 3D object recovery involves two primary stages [(Wang et al., 2 Feb 2026) §IV]:

1. Visible-Part Recovery: Solve a compressed sensing problem, using occlusion knowledge encoded via $V$ and a stack of RIS configurations $\{q_k\}$, to obtain estimated visible scattering/occupancy coefficients $\omega_v, \chi_v$ from $K$ wireless measurements.
2. Amodal Completion: With visible voxels $\chi_v$ recovered, a conditional diffusion model (built as a dual-branch U-Net) infers the complete, unoccluded shape occupancy $\chi$.

Generative Diffusion Model

The conditional diffusion model conducts shape completion by:

• Executing a forward diffusion process:

$$q(\chi_t|\chi_{t-1}) = \mathcal{N}\left(\sqrt{1-\beta_t}\chi_{t-1},\, \beta_t I\right), \quad t=1,\dots,T,$$

and a reverse denoising process parameterized by network weights $\theta$—with per-step predictions $\mu_\theta(\chi_t, t, \chi_v)$ for denoised occupancy [(Wang et al., 2 Feb 2026) §IV.C.2].

• Incorporating visible-shape features via a control branch in the decoder, guiding global-to-local completion.

Loss is mean squared error (MSE) between network-predicted and true noise: $$\mathcal{L}(\theta) = \mathbb{E}_{\chi_0, \epsilon, t} \big\|\epsilon - \epsilon_\theta(\chi_t, t, \chi_v)\big\|_2^2.$$

3. RIS Phase Shift Optimization

The accuracy of amodal reconstruction is highly sensitive to the sequence of RIS configurations used across $K$ measurements: $$\min_{q_1,\ldots,q_K}\; a \qquad \text{s.t. } \phi_{k,m} \in \mathcal{F},$$ where $a$ denotes the overall error metric [(Wang et al., 2 Feb 2026) §III.A].

Because the $a\leftrightarrow Q$ mapping (from RIS phase settings $Q$ to reconstruction error) lacks closed-form characterization, a supervised deep neural network (DNN) is trained to predict $a$ based on compact correlation features:

• Features: Global and local correlations $\{c_0, c_1, \ldots, c_L\}$, with $c_0$ measuring decorrelation of synthesized measurements ($R(QH_r)$) from the identity [(Wang et al., 2 Feb 2026) Eq. 11].
• Architecture: 7-layer fully connected, mapping $\mathbb{R}^{L+1} \rightarrow \mathbb{R}$.
• Phase Selection: Initialize $Q$ randomly, optimize via gradient descent with backpropagation through the DNN predictor, and quantize phases to their discrete set [Fig. 3, §III.B.3].

Alternative strategies for RIS configuration, such as codebook-based beam scanning or semidefinite relaxation (SDR), have been demonstrated to be less effective in the context of complex amodal recovery [(Kompostiotis et al., 28 Nov 2025) §2.3].

4. Sensing Algorithms and Experimental Validation

Algorithmic Strategies

The RIS configuration for wireless amodal sensing can follow:

• Beam Sweeping and Hierarchical Codebooks: Predefined phase pattern codebooks $\{\Theta(k)\}$ to scan spatial sectors, identifying reflective voxels or objects by correlation with strong received echoes [(Kompostiotis et al., 28 Nov 2025) §3.1–3.2].
• Compressed Sensing Recovery: Reconstruct sparse scattering profiles $x$ by minimizing $\|y-\Phi x\|_2^2 + \lambda \|x\|_1$ using measurements over $K$ RIS phase states [(Kompostiotis et al., 28 Nov 2025) §3.3].
• Learning-Based Inference: Mapping observed echo patterns $\{y_k\}$ to estimated object properties using DNNs, trained on labeled synthetic or experimental data [(Wang et al., 2 Feb 2026) §IV, (Kompostiotis et al., 28 Nov 2025) §3.4].

Empirical Results and Metrics

Experimental studies demonstrate:

• Spatial Resolution: Enhanced by the RIS aperture ($N$: number of elements, $b$: phase bit), with angular precision $\Delta\phi \approx \lambda / (N d\, \cos\theta)$ and range resolution $\Delta r \approx c/(2B)$ for bandwidth $B$ [(Kompostiotis et al., 28 Nov 2025) §4.1].
• Reconstruction Error: RIS-aided amodal sensing with optimized configurations achieves at least a 56.73% reduction in normalized L1 error vs. conventional schemes such as GAMP-Vision and MetaSketch on ShapeNet benchmarks ($N=10^3$, up to 65.54% lower error than vision-based amodal completion) [(Wang et al., 2 Feb 2026) §V.4.3].
• Experimental Prototype: 256-element 28 GHz RIS enables >90% detection probability ($P_D$), angular RMSE ≈2°, and range error ≈0.5 m over 8 m echo paths in non-line-of-sight (NLoS) scenarios, outperforming direct-path radar which fails in the same setup [(Kompostiotis et al., 28 Nov 2025) §5.3].

5. Performance Trade-Offs and Practical Considerations

Principal Trade-Offs

Parameter	Effect on Sensing	Associated Trade-Off
RIS aperture ($N$)	Finer spatial/angle resolution	Hardware cost, physical size
Phase resolution ($b$)	Reduces quantization loss	Complexity of control and electronics
#RIS configs ($K$)	Improves sampling density	Increased measurement time
Bandwidth ($B$)	Better range discrimination	Demands on RF chain, noise
ML model capacity	More accurate mapping	Need for labeled data and computation

Increasing $N$, $b$, or $K$ improves resolution and recovery accuracy but raises system complexity and acquisition time. Diminishing returns are observed as $K$ grows large. Mutual coupling, hardware impairments (e.g., phase error $\pm 5^\circ$), and computational demands of joint optimization must be managed [(Kompostiotis et al., 28 Nov 2025) §5.3] [(Wang et al., 2 Feb 2026) §V.4.3].

6. Limitations and Research Directions

Identified limitations include:

• Model Complexity: Training the generative diffusion model and DNN-based phase optimizer requires extensive offline computation and large datasets [(Wang et al., 2 Feb 2026) §V.5.1].
• Generalization: Reported pipelines are trained on synthetic (ShapeNet) objects; transfer to real-world, novel objects, or new materials is unverified.
• Scene Dynamics: Static scene assumptions preclude real-time tracking or operation under rapid occlusion changes.

Potential research avenues involve:

• Online Adaptation: Continual learning for error-predictor DNNs to adjust to evolving sensing environments.
• Multimodal Sensing: Fusing RIS-enabled wireless data with auxiliary vision or RGB-D measurements to reduce completion ambiguity.
• Real-Time Phase Optimization: Development of ultra-fast, possibly reinforcement-learning-driven RIS controllers for dynamic adaptation.
• Theoretical Analysis: Derivation of analytical or tractable bounds relating RIS phase patterns to amodal completion fidelity.
• Scalability: Extension to multi-object scenarios, distributed/cooperative RIS, and adaptation to mmWave/THz bands [(Wang et al., 2 Feb 2026) §V.5.2].

A plausible implication is that as RIS technology matures and real-time robust phase management becomes feasible, wireless amodal sensing will offer a high-privacy, robust alternative to conventional vision in occlusion-prone or privacy-sensitive environments.

7. Context and Related Paradigms

RIS-aided wireless amodal sensing complements and extends conventional wireless localization and radar by enabling virtual line-of-sight links, thus circumventing NLoS challenges. It differs fundamentally from passive propagation environments by turning the radio channel into an active sensing participant through programmable control (Kompostiotis et al., 28 Nov 2025). Key distinctions from vision-based amodal completion arise from the physically different nature of occlusions, privacy properties, and robustness in degraded visual environments (Wang et al., 2 Feb 2026).

Emerging prototypes and benchmarks validate the principle of RIS-aided sensing but underscore the importance of system-level co-design, including metrics such as detection probability ($P_D$), angular RMSE, and normalized reconstruction error across object categories and geometries. Ongoing work explores deploying larger-scale RIS, integrating with low-resolution and infrared vision, and lowering runtime computational complexity for in situ deployment.