Intelligent Sensing Layer in 6G Networks

• Intelligent Sensing Layer is a programmable and adaptive subsystem that collects multi-modal environmental data using RIS/LIS and near-sensor machine learning.
• It integrates physical-layer reconfigurability with advanced signal processing, cross-layer optimization, and edge AI for efficient 6G ISCC architectures.
• Practical implementations focus on optimizing sensing fidelity, communication throughput, and energy efficiency via joint resource management and dynamic reconfiguration.

An Intelligent Sensing Layer is a programmable, adaptive subsystem positioned at or near the physical edge of communication and computing infrastructures in next-generation wireless networks such as 6G. Its purpose is to acquire, extract, and optimize multi-modal environmental information—specifically angle, range, Doppler, and other salient features—by leveraging physical-layer reconfigurability (with elements such as RIS/LIS), in-situ or near-sensor machine learning, and cross-layer optimization. This layer feeds high-quality, context-rich sensing data into higher layers for communication, computation, resource management, and autonomous decision-making, while allowing dynamic reconfiguration to maintain a joint optimum of sensing fidelity, communication throughput, and energy/resource efficiency (Li et al., 2024, Chepuri et al., 2022, Liu et al., 30 Apr 2025, Vaca-Rubio et al., 2020, Wymeersch et al., 16 May 2025).

1. Functional Architecture and Cross-Layer Roles

The Intelligent Sensing Layer (ISL) is most commonly instantiated as the lowest logical layer in ISCC (Integrated Sensing, Communication, and Computation) frameworks, typically realized via a network of IoT/edge devices equipped with active sensing frontends (e.g., mmWave/THz radar, RF, optical), and one or more programmable electromagnetic surfaces—Reconfigurable Intelligent Surfaces (RIS) or Large Intelligent Surfaces (LIS) (Li et al., 2024, Vaca-Rubio et al., 2020, Liu et al., 2022). Its key functions include:

• Emitting probing signals and collecting reflections (direct and surface-assisted).
• Performing parameter estimation (angle, range, Doppler) from multi-path echoes.
• Feeding sensing metrics (e.g., Cramér–Rao Bound, detection probability) to resource managers and higher-layer controllers.
• Triggering real-time reconfiguration of surface elements to steer beams, adapt sensing/communication tradeoffs, and maintain task offloading quality (Li et al., 2024, Wymeersch et al., 16 May 2025, Cui et al., 2024).
• Supporting data-intensive ML pipelines (e.g., feature extraction, anomaly/event detection) local to the edge node or ISL.

A typical ISL is equipped with a control plane enabling high-frequency adaptivity, interfacing downward to the physical channel and upward to edge computing/central control, as outlined in Table 1.

Layer	Function	Reconfigurable Elements
Sensing Layer (ISL)	Sensing, feature extraction	RIS/LIS, edge ML, active sensors
Communication	Uplink/offload data transfer	RIS beamforming, channel scheduling
Computation	Edge/cloud inference/decision	Virtualized servers, model partition

2. Physical and Mathematical Models

The ISL tightly integrates physical-layer reconfigurability with cross-layer signal processing:

• Signal Model: The received communication and radar echoes at an edge node are modeled as combinations of direct and RIS-reflected channels (Li et al., 2024):
  • Direct channel: $h_d \in \mathbb{C}^{N \times 1}$
  • RIS-reflected: $F \in \mathbb{C}^{M \times 1}$ (IoT → RIS), $G \in \mathbb{C}^{N \times M}$ (RIS → edge)
  • Equivalent channel: $h_{eq}^H = h_d^H + F^H \Theta G$, with $$\Theta = \operatorname{diag}(e^{j\phi_1},...,e^{j\phi_M})$$
• Radar Echo Processing: Monostatic echo at device:

$$y_r = \beta\,a_r^H(\theta)\left[h_d + G^H \Theta F\right] a_t(\theta) s + w_r$$

• Parameter Estimation: Maximum Likelihood and Cramér–Rao Bound analysis are used for angular and range resolution, with multi-sector/multi-panel geometries adopted for full-space coverage and to eliminate angle blind spots (Zhang et al., 2024).
• Computation Offloading: The ISL serves as an offload-computation decision point, minimizing end-to-end latency under constraints:

$t_i^{sense} + t_i^{off} + t_i^{ec} \leq T_i^{max}$

• Joint Optimization: Cross-layer optimization is common:

$$\max_{\{w_i, \Theta, x_i, f_i\}}\ \alpha \sum_i R_i + \beta \sum_i P(\theta_i) - \gamma \sum_i E_i$$

Subject to power, phase, latency, and CPU constraints (Li et al., 2024, Liu et al., 30 Apr 2025).

3. Intelligent Reconfigurable Surfaces and Time-Modulated Architectures

Modern ISLs increasingly exploit advanced programmable metasurfaces:

• RIS/LIS: Passive panels of sub-wavelength elements with variable phase (and sometimes amplitude), imparting programmable, fine-grained steering and focusing properties. Used for both echo enhancement and multipath environment control.
• Multi-sector IS: Panels arranged in geometric prisms yield 360° coverage and eliminate directional “blind spots” in sensing, with proven gain in probing power and angle sensitivity, further enhanced by sectorally applied directive antennas (Zhang et al., 2024).
• Time-Modulated IRS (TM-IRS): Surfaces in which each element is rapidly switched (on-off or between quantized phases) during a symbol period, creating additional temporal degrees of freedom (harmonic beampattern engineering). Machine learning, specifically GFlowNets (generative flow networks), enable stochastic design in the high-dimensional configuration space, optimizing for joint sensing, communication, and secrecy objectives—e.g., maximizing legitimate user rate while scrambling eavesdropper directions (Tao et al., 6 Sep 2025).
• Sensing-communication Power Allocation: Optimization of RIS placement and joint power allocation per surface achieves near-optimal coverage and communication/sensing tradeoff, with analytical expressions for coverage, spectral efficiency, and the impact of phase estimation errors (Luo et al., 2024).

4. Edge AI, ML, and Feature Compression

The ISL integrates edge AI both for signal processing and as a key resource allocator:

• Near-sensor machine learning: Tiny-AI modules directly coupled to sensors (cameras, radar) make real-time on-board decisions (e.g., objectness detection, anomaly flagging), drastically reducing upstream data volume by transmitting only “valuable” data (Huang et al., 2024).
• Feature compression: Intermediate deep feature compression allows an ISL to offload compact, generalizable feature representations (from the first $L$ layers of a network) rather than raw sensor data or top-layer-specific features, balancing front-end and network compute load with offloading/storage cost. Compression exploits spatial transformation, sparsity and standard codecs, with further gains available by introducing standardized bitstream syntax (Chen et al., 2018).
• Distributed collaborative inference: Hybrid edge–cloud partitioning of DNNs enables K-tier ISL architectures to adaptively splice local, MEC, and remote cloud inference, with ISAC devices offloading intermediate features optimizing for joint sensing beam, partition layer, and compute allocations (Liu et al., 30 Apr 2025).

5. Protocols, Algorithms, and Cross-Layer Optimization

ISL operation mandates advanced algorithms for joint optimization and control:

• Alternating Optimization (AO): Alternately optimize RIS phases, beamformers, and offloading assignments, using convex solvers for w, SDR/manifold optimization for RIS phases, all subject to tight cross-layer constraints (Li et al., 2024, Liu et al., 2022, Chepuri et al., 2022).
• AI/ML-based control: Deep Reinforcement Learning (DDPG, PPO), GFlowNet policy learning, and cross-entropy methods are used for high-dimensional configuration and partitioning (e.g., selecting partition points, surface parameterizations, dynamic adaptation to channel variations) (Tao et al., 6 Sep 2025, Li et al., 2024, Liu et al., 30 Apr 2025).
• Sensing–communication–computation co-design: Multi-objective cost functions balance CRB, sum rate, and energy, and are implemented by resource optimization modules at the ISL, informed by real-time feedback on sensing and channel metrics (Wymeersch et al., 16 May 2025, Cui et al., 2024).
• Feedback and orchestration: The ISL supports protocol steps of probing, decision, reconfiguration, offloading, and feedback, efficiently communicated between devices, RISs, and edge servers. Real-time performance relies on fast channel estimation (AI-driven unfolding, compressive sensing), robust hardware design under quantization, and predictive adaptation under mobility (Li et al., 2024).

6. Performance Metrics, Benchmarks, and Experimental Results

Empirical evaluation and analytical benchmarks established in the literature include:

• Sensing Performance: Detection probability ($P_d$), CRB of angle/range, beampattern gain $P(\theta_0)$, and radar SNR. Full-space, worst-case-minimized CRB can be achieved with multi-sector architectures and directive patterns (Zhang et al., 2024).
• Communication Performance: Sum-rate, uplink/offloading rate, energy per bit, and spectral efficiency. RIS-empowered ISLs report uplink rate gains of up to 30% for challenging cell-edge cases at moderate numbers of RIS elements ($M=40-60$) (Li et al., 2024).
• Computation and Latency: End-to-end latency, minimal inference time under task partitioning, and compute energy. Joint ISL partitioning and cross-layer optimization reduce overall completion time by up to 40% compared to two-tier baselines (Liu et al., 30 Apr 2025).
• Energy and Bandwidth Efficiency: Deployment of near-sensor ML and selective transmission mechanisms results in >85% combined storage, bandwidth, and energy savings, with negligible (<5%) impact on detection completeness (Huang et al., 2024).
• Robustness and Adaptivity: AI-driven ISLs adapt to environmental changes, maintain performance under SNR degradation, and are resilient to model mismatch in practical hardware (Vaca-Rubio et al., 2020, Li et al., 2024).
• Benchmark Pipelines: Classical statistical (GLRT) and modern end-to-end (denoising autoencoder, deep feature transfer) pipelines are established for radio imaging and anomaly detection (Vaca-Rubio et al., 2020, Vaca-Rubio et al., 2021).

7. Challenges, Practical Considerations, and Future Directions

Practical deployment of ISLs faces several domain-specific challenges:

• Channel Estimation Overhead: Cascaded channel estimation scales linearly or superlinearly with surface size, requiring new compressive, on-off pilot, or AI-accelerated methods (Li et al., 2024, Liu et al., 2022).
• Hardware Impairments: Finite phase-resolution, amplitude–phase coupling, mutual coupling, and frequency dependency require robust or stochastic design approaches.
• Mobility and Orchestration: Fast adaptation of RIS/ISL control under user and environmental mobility is critical. Predictive and learning-based beamforming, as well as digital twin frameworks, are emerging solutions for dynamic reconfiguration and service continuity (Li et al., 2024).
• Edge Intelligence Integration: ISLs increasingly blend edge ML (e.g., federated learning, knowledge distillation, lightweight neural architectures), digital twins, and cross-layer orchestration for rapid adaptation, privacy, and resource efficiency (Cui et al., 2024, Wymeersch et al., 16 May 2025).
• Security and Privacy: Physical-layer security via TM-IRS, privacy-preserving feature processing, and privacy-aware federated training are active topics (Tao et al., 6 Sep 2025, Chepuri et al., 2022).
• Scalability and Real-Time Control: Ultra-dense and large-scale deployments mandate scalable, low-latency control protocols and surface clustering/group updates (Zhang et al., 2022).
• New Application Scenarios: UAV-RIS, road-side/vehicular RIS, and distributed multi-RIS in urban environments, along with integration of digital twins for predictive or intent-aware configuration, are highlighted as next strategic steps (Li et al., 2024, Cui et al., 2024, Wymeersch et al., 16 May 2025).

Intelligent Sensing Layers in 6G ISCC and ISAC architectures thus establish an adaptive, programmable substrate for extracting, optimizing, and routing environmental and context data throughout edge, communication, and computation layers. Their implementation draws on cross-disciplinary advances in physical-layer metasurfaces, statistical and ML signal processing, systems optimization, and protocols, enabling robust, efficient, and context-rich operation at the physical edge (Li et al., 2024, Wymeersch et al., 16 May 2025, Liu et al., 30 Apr 2025, Zhang et al., 2024, Tao et al., 6 Sep 2025).