Distributed ISAC: Architectures & Metrics

• Distributed Integrated Sensing and Communication is a network of geographically dispersed nodes that cooperatively perform both sensing and communication tasks.
• The system jointly optimizes metrics like positioning accuracy, spectral efficiency, and energy use while balancing power and bandwidth trade-offs.
• Architectures span decentralized to centralized fusion with explicit control-data plane separation, enabled by semantic signal processing and advanced synchronization.

Distributed Integrated Sensing and Communication (D-ISAC) denotes the extension of integrated sensing and communication from a single transceiver or centralized node to a coordinated network of geographically separated transmitters, receivers, reconfigurable surfaces, user devices, and edge/cloud processors that jointly illuminate, observe, communicate, and fuse information over space and time. In 6G-oriented formulations, D-ISAC is often coupled to intelligence, semantics, and goal-oriented control, so that the system moves from raw-data fusion and Shannon-centric optimization toward semantic composition and task-level KPIs such as positioning accuracy, orientation accuracy, velocity estimation, spectral efficiency, and energy efficiency (Strinati et al., 2024, Strinati et al., 2024, Han et al., 2 May 2025).

1. Conceptual scope and taxonomy

D-ISAC extends ISAC from a single BS or co-located array to multi-node architectures in which distributed nodes cooperatively serve communication users and sense targets through multi-static geometries, local processing, and fusion (Zeng et al., 2023, Strinati et al., 2024). At the network level, the literature distinguishes non-cooperative, interference-management, decentralized, and centralized cooperation regimes, with increasing CSI or raw-data exchange, signaling overhead, and achievable sensing-and-communication gains (Han et al., 2 May 2025).

A recurrent distinction is between Distributed ISAC (D-ISAC) and Distributed Intelligent ISAC (DISAC). D-ISAC emphasizes the distributed architecture itself: multiple sensing transmit nodes, sensing receive nodes, passive sensing nodes, RRHs, APs, RISs, and UEs cooperate over extended space-time. DISAC adds AI/ML, semantic-native communication, and goal-oriented operation, so that sensing activation, waveform shaping, resource allocation, and knowledge exchange are conditioned on task relevance rather than only on bit transport or raw-data aggregation (Strinati et al., 2024, Strinati et al., 2024).

Architecturally, D-ISAC appears in distributed MISO, SIMO, and MIMO forms, as well as in TX/RX-separated half-duplex and collocated full-duplex realizations (Han et al., 2 May 2025). In the multisensor view, it is effectively the distributed MIMO radar counterpart of multi-user MIMO communication, reusing cellular access resources for scene illumination, sensing, data transportation, computation, and fusion (Thomä et al., 17 Nov 2025). A common misconception is that D-ISAC is merely “more nodes” added to conventional ISAC. The surveyed work instead treats it as a network service with distributed apertures, multi-link access, coordinated fusion, and explicit control-plane machinery for synchronization, scheduling, handover, and semantic exchange (Strinati et al., 2024, Stylianopoulos et al., 17 Apr 2025).

2. Architectural principles and functional decomposition

The node set in D-ISAC is heterogeneous. Across the surveyed architectures it includes base stations or transmission reception points, RRHs, distributed APs, STNs, SRNs, PSNs, RISs, UAVs, vehicles, sidelink-capable UEs, and non-3GPP sensors such as cameras, lidar, radar, traffic stations, weather stations, and GNSS receivers (Strinati et al., 2024, Strinati et al., 2024). In D-MIMO and cell-free variants, distributed APs or radio stripes form a logical array under a central unit or edge controller, enabling coherent or non-coherent joint transmission, multi-static sensing, and cooperative localization (Guo et al., 2024, Rosabal et al., 10 Apr 2026).

The control and data planes are explicitly separated. The control plane carries semantic instructions, synchronization and calibration signaling, function orchestration across SRN groups, and identifiers for passive objects; the data plane transports raw IQ, extracted channel parameters, feature-level summaries, or semantic artifacts depending on node capability and task needs (Strinati et al., 2024). Recent DISAC architectures further introduce a Sensing Processing Function (SePF) at network nodes and a Sensing Management Function (SeMF) for orchestration, scheduling, data fusion, semantic policy enforcement, and service exposure, including O-RAN-aligned deployment options (Stylianopoulos et al., 17 Apr 2025).

Data representation is multi-level. The surveyed works enumerate raw IQ, extracted channel parameters, Range–Angle–Doppler tensors, point clouds, parametric object models, tracks, maps, deep embeddings, and semantic summaries as possible exchange units (Stylianopoulos et al., 17 Apr 2025, Strinati et al., 2024). This layered representation is central because D-ISAC is constrained not only by radio resources but also by fronthaul, backhaul, and compute budgets. Hierarchical processing therefore places local data transformation and compression near the sensing nodes, with regional or global fusion at edge or cloud sites (Strinati et al., 2024, Guo et al., 2024).

Several concrete infrastructures instantiate these principles. In distributed antenna networks, a central processor coordinates RRHs and partitions each frame into orthogonal communication and sensing phases under fronthaul and QoS constraints (Xu et al., 2022). In radio stripe systems, homogeneous antenna processing units share a common bus for synchronization, data transfer, and power supply, and each unit can be dynamically assigned to communication or sensing mode (Rosabal et al., 10 Apr 2026). In RDARS-assisted systems, programmable elements switch between reflection mode and connected mode, so that one substrate simultaneously acts as a reconfigurable surface and as a distributed receive aperture (Wang et al., 2023).

3. Signal models, metrics, and joint design objectives

Most D-ISAC formulations reuse standard communication and sensing models but distribute them across nodes. A canonical communication model writes

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n},$

while a canonical sensing echo model writes

$$s(t) = \sum_k \alpha_k x(t-\tau_k)e^{j 2\pi f_{d,k} t} + w(t),$$

with delays, Dopplers, path gains, and additive noise defined per target or path (Strinati et al., 2024). In distributed MIMO-OFDM settings, the sensed response is extended to multiple TX/RX sets, enabling complementary geometries, larger effective apertures, and improved GDOP (Strinati et al., 2024).

The dominant sensing metrics are Fisher information, Cramér–Rao lower bounds, detection probability, false alarm probability when explicitly modeled, RMSE, and application-specific localization or tracking errors. Communication is measured through per-user SINR, rate, spectral efficiency, latency, reliability, and energy efficiency. D-ISAC formulations often combine these in a weighted objective such as

$$\max_{W,x,\Phi}\ \lambda_c R_{\mathrm{comm}}(W,x,\Phi) + \lambda_s U_{\mathrm{sense}}(W,x,\Phi)$$

subject to power, latency, compute, detection, and CRB constraints (Strinati et al., 2024). Network-level analyses further use communication and sensing coverage probability, area spectral efficiency, potential spectral efficiency, and joint ISAC ergodic rate (Han et al., 2 May 2025).

Resource coupling is fundamental. In stochastic-geometry analysis for distributed peer-to-peer ISAC, sensing and communication share total power and bandwidth budgets, and the resulting trade-off is quantified through the probability of detection and the communication coverage probability. The derived results indicate that communication is more sensitive to bandwidth, while sensing is more sensitive to power, leading to the design guideline of allocating a larger fraction of bandwidth to communication and a larger fraction of power to sensing, especially in dense deployments (Li et al., 2023).

Waveform and signaling design differ sharply across architectures. In distributed antenna networks coordinated by a central processor, one energy-focused solution is to separate the ISAC frame into a communication phase and a sensing phase, with the sensing phase using short-duration pulses and the communication phase using low-power long-duration signaling (Xu et al., 2022). In noncoherent OFDM D-ISAC, multiple ISAC nodes cooperate without phase-level synchronization; the communication SINR is modeled under noncoherent coordinated multi-point transmission, and the sensing design optimizes per-subcarrier signals because time-of-flight estimation quality depends directly on subcarrier-domain structure (Han et al., 30 Jan 2025). A useful corrective to a common misconception is that distributed sensing always requires phase-coherent cooperation. Noncoherent D-ISAC explicitly avoids phase-level synchronization, while accepting that useful signals and interference sum in power rather than as aligned complex amplitudes (Han et al., 30 Jan 2025).

4. Cooperation, fusion, synchronization, and continuity management

Distributed cooperation appears in three main fusion regimes: centralized, hierarchical, and fully distributed. Centralized schemes minimize local processing and push sensing outputs or raw observations to a fusion center; hierarchical schemes perform local fusion and compression before regional or global aggregation; fully distributed schemes rely on peer-to-peer exchanges, graph consensus, or diffusion updates among neighboring nodes (Strinati et al., 2024, Li et al., 2024). Leaderless cooperative sensing has been demonstrated in networked ISAC through a two-step distributed total least squares algorithm with adapt-then-combine diffusion, block sparsity, and low-dimensional intermediate estimate exchange, explicitly to avoid dependence on a centralized leader node (Li et al., 2024).

Tracking continuity is a specialized D-ISAC problem. In monostatic multi-BS DISAC, target trajectories must be handed over across limited fields of view. A random-finite-set solution based on a trajectory Poisson multi-Bernoulli mixture filter shares only essential trajectory information, encoded as a handover PPP intensity,

$$\lambda_{H,k+1|k}^{(m,j,i)}(X)=w_{k+1|k}^{(m,j)} r_{k+1|k}^{(m,j,i)} f_{k+1|k}^{(m,j,i)}(X),$$

which the receiving BS adds to its predicted PPP intensity before the local update (Ge et al., 2024). In cooperative user tracking for distributed MIMO, a global GM-PHD filter combined with a field-of-view-aware AP scheduling policy achieves accurate 3D tracking while reducing the number of active APs (Xu et al., 5 Nov 2025).

Synchronization is the most persistent systems bottleneck. Time, frequency, and phase alignment across STNs, SRNs, APs, RRHs, and RISs directly affect delay, Doppler, angle estimation, coherent combining, and the achievable CRB (Strinati et al., 2024, Guo et al., 2024). Network-level surveys classify over-the-air synchronization into cross-antenna, path-resolved, fingerprint-spectrum, and cooperation-based methods, and further exploit offset reciprocity in reciprocal bi-static links,

$$\Delta_{n,m}^{t}=-\Delta_{m,n}^{t},\qquad \Delta_{n,m}^{f}=-\Delta_{m,n}^{f},$$

to enable super-resolution time-offset and CFO estimation (Han et al., 2 May 2025). In D-MIMO deployments, coherent FR1 indoor positioning can yield orders-of-magnitude PEB gains relative to conventional time-coherent positioning, but those gains are contingent on stringent synchronization and calibration (Guo et al., 2024).

Another misconception is that every AP must remain active to preserve tracking quality. Real measurement-based cooperative tracking shows otherwise: centimeter-level performance can be maintained with FoV-aware activation, indicating that AP management is an intrinsic part of D-ISAC design rather than a secondary implementation detail (Xu et al., 5 Nov 2025).

5. Semantic-native and learning-driven extensions

The intelligence layer in DISAC is not an add-on to physical-layer cooperation; it restructures what is sensed, what is transmitted, and how fusion is performed. Semantic communication is defined as transmitting meaning or knowledge rather than raw bits, using a semantic channel and semantic encoders/decoders. The corresponding semantic plane implements semantic extraction, semantic composition, and semantic instruction, so that sensing, waveform design, and resource allocation are activated according to application goals and the value of information (Strinati et al., 2024, Strinati et al., 2024).

This shift changes the optimization target. Instead of optimizing only Shannon metrics, DISAC introduces task-level KPIs and canonical semantic metrics such as semantic distortion

$D_{\mathrm{sem}} = E[d(M,\hat M)],$

task success probability

$P_s = P(\text{task outcome} \ge \text{threshold}),$

and mutual information of meaning $I(M;\hat M)$ (Strinati et al., 2024). The practical consequence is a move from raw-data fusion toward semantic composition of tracks, maps, relations, and context-rich descriptors, with high-level artifacts preferred over raw sensing streams when communication capacity is limited (Strinati et al., 2024).

AI modules span federated learning, distributed auto-encoders, graph signal processing, multi-agent learning, on-device inference, and edge/cloud training (Strinati et al., 2024). In distributed FMCW-based vertical federated edge learning, multiple edge devices sense a common target, transform local data into intermediate vectors, and exchange those vectors rather than raw sensing data for collaborative recognition; this is a concrete D-ISAC realization because the same waveform and hardware support both sensing and feature exchange (Liu et al., 2022). Privacy and security are treated primarily through reduced raw-data exposure, flexible data models, and privacy-preserving distributed learning rather than through mandatory centralized data collection (Strinati et al., 2024, Stylianopoulos et al., 17 Apr 2025).

A plausible implication is that semantic-native D-ISAC reduces not only payload size but also orchestration complexity, because network decisions can be conditioned on compact, task-aligned state descriptions instead of large heterogeneous measurement sets. The surveyed papers stop short of claiming a universal semantic standard, and several identify semantic KPI definition, semantic channel modeling, and interoperability as unresolved research problems (Strinati et al., 2024, Stylianopoulos et al., 17 Apr 2025).

6. Applications, prototypes, quantitative benchmarks, and open problems

D-ISAC is motivated by extended environments in which centralized ISAC is coverage-limited, backhaul-limited, or geometrically weak. Representative scenarios include intelligent transportation and urban digital twins, smart factories and logistics, public safety and intrusion detection, XR, human activity recognition, healthcare-oriented privacy-preserving sensing, environmental monitoring, VRU protection at smart intersections, and secure fiber-based sensing-and-communication networks (Strinati et al., 2024, Stylianopoulos et al., 17 Apr 2025, Xu et al., 2024).

Several papers provide concrete quantitative targets or experimental benchmarks.

Benchmark	Value	Source
Spectral efficiency target	+30% over 3GPP R18 MIMO	(Strinati et al., 2024)
User positioning accuracy target	0.01 m (FR2), 0.1 m (FR1)	(Strinati et al., 2024)
Energy efficiency target	+40% improvement in FR2 over 3GPP TR 38.684 baseline	(Strinati et al., 2024)
User orientation accuracy target	<1° in FR2	(Strinati et al., 2024)
Object positioning accuracy target	<1 m (FR1), 0.1 m (FR2)	(Strinati et al., 2024)
Object velocity estimation target	<1 m/s (FR1)	(Strinati et al., 2024)
Quantum-fiber communication	approximately 0.7 Mbits/s SKR per user at 10 km; network capacity 8	(Xu et al., 2024)
Quantum-fiber sensing	1 Hz to 2 kHz; 0.50 nε/$\sqrt{\rm Hz}$; 0.20 m	(Xu et al., 2024)
RDARS prototype	81.8 Mbit/s over 20 MHz; within 5° of ground truth	(Wang et al., 2023)
Cooperative tracking	average trajectory RMSE ≈ 0.09 m; average active panels ≈ 4.78 out of 8	(Xu et al., 5 Nov 2025)
Vertical FEEL realization	up to 98% recognition accuracy; up to 8% improvement over benchmarks	(Liu et al., 2022)
Radio stripe reconfiguration	average rate degradation of ≈0.8773 Mbit/s from best to worst S/C configuration	(Rosabal et al., 10 Apr 2026)

Prototype diversity is notable. Wireless demonstrations include distributed sounders with FR1/FR2, virtual XL-MIMO, DMIMO, RIS, distributed auto-encoders, graph signal processing, 5G-NR waveform testing at 27 GHz, semantic-aware waveform platforms, backscatter and sensing RIS in S/Ka bands, and environmental sensing via radio-SLAM (Strinati et al., 2024). RDARS shows that a reconfigurable distributed antenna-and-surface platform can localize a user without compromising uplink rate (Wang et al., 2023). Radio stripe systems show that increasing the number of devices and sensing APUs boosts sensing precision at the expense of degrading sum rate, while the sum rate remains constant for a given number of communication APUs regardless of their positions under the paper’s homogeneous path-loss and MRT assumptions (Rosabal et al., 10 Apr 2026). Optical-fiber D-ISAC extends the paradigm into the quantum domain through an integrated sensing and quantum network that combines multi-user CV-QKD with distributed vibration sensing in a star topology (Xu et al., 2024).

The main open problems recur across otherwise different implementations. The list includes synchronization at scale, heterogeneous hardware and calibration, semantic metrics and channel models, joint radio/sensing/compute scheduling, scalable high-resolution processing in rich multipath, object identifiers and inter-cell handover, RIS/D-MIMO/XL-MIMO deployment and control, and standardization gaps for sensing-native network functions and semantic-native signaling (Strinati et al., 2024, Han et al., 2 May 2025). Another recurring lesson is that performance does not improve monotonically with every architectural degree of freedom: in radio stripes, changing the number of antennas can have a non-monotonic effect on sensing because array gain and illumination uniformity pull in opposite directions (Rosabal et al., 10 Apr 2026). D-ISAC therefore remains a fundamentally cross-layer design problem in which geometry, synchronization, fusion level, waveform structure, semantics, and compute placement are all first-order variables rather than secondary implementation choices.