Integrated Sensing, Communication, and Computing

• ISCC is a unified system design integrating sensing, communication, and computing to optimize end-to-end task performance and resource efficiency.
• It shifts the design focus from isolated metrics like data rate or local compute speed to joint outcomes such as inference accuracy, latency, and energy consumption.
• Applications include vehicular platforms, edge-AI systems, and industrial IoT, where shared time–space–frequency–computing resources are allocated dynamically for optimal performance.

Integrated Sensing, Communication, and Computing (ISCC) denotes a unified design of wireless sensing, data transmission, and computation in which the three functions are treated as parts of a single task-oriented system rather than as separate subsystems. In the 6G and edge-intelligence literature, this unification is motivated by applications whose performance is determined jointly by sensing quality, communication efficiency, and computation capability, so the relevant objectives shift from rate- or BER-centric design toward task-level quantities such as inference accuracy, learning convergence, latency, energy consumption, and, in some formulations, information gain per unit resource (Zhu et al., 2022, Chen et al., 2024, Xing et al., 2023). The term is used both as a broad architectural paradigm and as a family of concrete system models, ranging from information-oriented vehicular platforms based on shared time–space–frequency–computing resources to wireless learning systems that couple online data acquisition, local model updates, and over-the-air aggregation (Chen et al., 2024, Liang et al., 2024).

1. Conceptual foundations

ISCC is commonly presented as an extension of Integrated Sensing and Communication (ISAC) that elevates computing to a first-class design dimension. In this view, sensing acquires data from the physical or digital environment, communication transports raw data or task-relevant representations, and computation performs local, edge, or cloud processing for training, inference, estimation, or control; the key departure from conventional designs is that these functions are optimized jointly and with awareness of the downstream task (Zhu et al., 2022). Several works describe this as a task-oriented or information-oriented shift: instead of separately optimizing detection probability, data rate, and computational complexity, the system is designed to maximize end-to-end task utility under coupled resource constraints (Chen et al., 2024, Xing et al., 2023).

The literature also emphasizes that ISCC is not merely a slogan for 6G convergence. One line of work defines it as a “three-dimensional cooperation” of sensing, communication, and computing built on a common time–space–frequency–computing resource pool, explicitly decoupled from any single function and reallocated according to information demand (Chen et al., 2024). Another line frames it as the task-driven “outer envelope” of pairwise integrations—ISAC for sensing plus communication, AirComp-style integrated computation and communication, and neuromorphic sensing-computation pipelines—arguing that full integration arises when all three are co-designed around the same edge-AI objective (Xing et al., 2023).

A recurrent theme is the inadequacy of separated design. Conventional systems may maximize raw sensing fidelity, communication throughput, or local compute efficiency in isolation, yet still fail to meet end-to-end requirements of autonomous driving, smart factories, or XR. The ISCC literature therefore treats sensing time and quality, communication bandwidth and power, and computation placement and speed as jointly coupled variables whose value is defined by downstream task performance rather than by subsystem-local metrics (Zhu et al., 2022).

2. Architectural patterns and shared resources

A generic ISCC architecture comprises edge devices that sense and perform local preprocessing, edge servers that aggregate information and execute heavier workloads, cloud data centers for large-scale coordination, and wireless infrastructure that supports both data exchange and, in some cases, sensing itself (Zhu et al., 2022). This architecture appears in multiple variants. In information-oriented IoAV systems, sensing data about non-connected targets are generated by connected automatic vehicles, offloaded via mmWave links to road-side units, and processed at local sensing and computing units or at edge sensing and computing units; the associated resource model is expressed through a Twin Resource Pool consisting of time–space, time–frequency, and time–computing planes (Chen et al., 2024). In mobile-edge settings, three-tier Cloud–Edge–Terminal architectures allow terminals to process sensing tasks locally or offload them to MEC servers or a cloud server while reusing ISAC waveforms for both data transfer and target probing (Liu et al., 2024, Liu et al., 2024).

These architectures are coupled by explicit latency and energy models. In ISCC-enabled Air-FEEL, each device incurs sensing time $T_{k,s}^{(t)} = b_k^{(t)}\tau_s$, computation time $T_{k,c}^{(t)} = b_k^{(t)}C/f_k$, and uplink AirComp time $T_u^{(t)} = \lceil N/M\rceil \tau_u$, with corresponding sensing, computing, and communication energy terms that all depend on batch size, sensing power, CPU frequency, and aggregation parameters (Wen et al., 21 Aug 2025). In three-tier offloading formulations, latency combines local execution, uplink transmission, MEC processing, and, when used, MEC–cloud backhaul, while power budgets couple beamforming, local CPU frequency, and sensing requirements (Liu et al., 2024, Liu et al., 2024). This suggests that “architecture” in ISCC is not only a placement problem; it is also a statement about which resources are fungible across sensing, communication, and computation and which are not.

A further architectural trend introduces spatial reconfigurability. IRA-enabled ISCC systems place multiple intelligent rotatable antennas and an edge server at a triple-functional base station so that the same infrastructure can collect sensing echoes, serve communication users, and support task offloading while adapting antenna boresight directions to task geometry (Xiong et al., 16 Jun 2025). Here the architectural novelty lies less in adding another tier than in exposing antenna orientation itself as an ISCC control variable.

3. Signal-level integration and learning workflows

At the signal-processing level, ISCC often relies on mechanisms that collapse communication and computation into the physical layer. Over-the-air computation (AirComp) is central in this respect: multiple devices simultaneously transmit analog symbols, and the receiver exploits waveform superposition to recover a function such as a sum, average, feature aggregation, or model update rather than decoding each stream separately (Zhu et al., 2022). In federated edge learning, this supports Air-FEEL and related designs in which the communication channel itself performs part of the aggregation required by learning (Wen et al., 21 Aug 2025).

A canonical realization appears in federated learning with integrated sensing, communication, and computation. In that framework, each communication round begins with global-model broadcast, continues with online sample sensing at edge devices, performs local training on cumulative sensed data, and ends with over-the-air aggregation of local models or gradients. The server attempts to recover the weighted sum

$$\tilde{\mathbf{u}}_t=\sum_{n=1}^N \rho_t^n \mathbf{u}_t^n,\qquad \rho_t^n=\frac{S_t^n}{S_t},$$

with communication mismatch represented by an explicit error term $\boldsymbol{\varepsilon}_t$ (Liang et al., 2024). Two concrete algorithms are analyzed in this setting: FedAVG-ISCC, which uses multiple local SGD steps and uploads local models, and FedSGD-ISCC, which computes full-batch gradients on cumulative sensed data and uploads gradients (Liang et al., 2024).

Theoretical results show that sensing and communication errors enter the convergence bounds explicitly. Sample collection affects the gradient through the mixture of old cumulative data and newly sensed data, while communication noise propagates through the analog aggregation stage (Liang et al., 2024). Under IID data, FedAVG-ISCC benefits from multiple local steps and outperforms FedSGD-ISCC; under non-IID data, the same local-update mechanism amplifies drift, so FedSGD-ISCC is more robust. FedSGD-ISCC is also more resilient to communication errors because its error term is not amplified by the multiple-local-step factor that appears in FedAVG-ISCC (Liang et al., 2024).

A closely related Air-FEEL formulation integrates FMCW radar sensing, local stochastic-gradient computation, and one-shot AirComp aggregation. Each sensing period generates one training sample through sampling, SVD-based clutter cancellation, STFT, vectorization, and normalization; the resulting sensed samples are then used for recognition-model training, and the aggregated gradient is corrupted by both sensing noise and AirComp distortion (Wen et al., 21 Aug 2025). The resulting convergence analysis separates three variance sources: stochastic gradient noise, sensing-induced gradient noise, and AirComp-induced noise. The paper further states that this is the first mathematical characterization of how wireless sensing quality, including residual clutter and sensing SNR, affects FEEL convergence (Wen et al., 21 Aug 2025).

ISCC has also been formulated directly over OFDM physical layers. In uplink OFDM ISCC with simultaneous target sensing and AirComp, the objective is to minimize computational MSE by jointly optimizing the transmitting vector and aggregation vector, while sensing performance is constrained through CRLB-based requirements for target distance and velocity estimation (Dong et al., 7 Mar 2025). This illustrates a general pattern: the same waveform parameters that control sensing fidelity also determine the accuracy of over-the-air function computation.

4. Optimization paradigms and representative methods

Because ISCC couples variables across sensing, communication, and computation, its optimization problems are usually non-convex, mixed-integer, or both. A generic task-oriented template maximizes a task metric $$\mathcal{A}(\boldsymbol{\theta}_{\text{s}},\boldsymbol{\theta}_{\text{c}},\boldsymbol{\theta}_{\text{comp}})$$ subject to latency, energy, and protocol constraints, with sensing, communication, and computation parameters optimized jointly rather than sequentially (Zhu et al., 2022). Within that template, several distinct paradigms have emerged.

Paradigm	Representative objective	Representative method
Task-oriented edge-AI design	Accuracy or convergence under latency/energy limits	Joint optimization of sensing, communication, and computation variables (Zhu et al., 2022)
Information-oriented IoAV design	Net profit $\Pi = R_{\text{info}} - C_{\text{TSFC}}$	IRTP with A2GNN, distributor and purchaser workers (Chen et al., 2024)
ISCC-enabled Air-FEEL	Minimize variance-type convergence term	Alternating optimization of batch size and resource allocation (Wen et al., 21 Aug 2025)
Three-tier offloading and beamforming	Minimize sensing-task execution latency	ADMM, fractional programming, WMMSE, and SCA (Liu et al., 2024)
Multi-device action-recognition ISCC	Maximize sensing accuracy under power, delay, and edge-CPU constraints	ADMM-based distributed algorithm with closed-form edge CPU allocation (Chen et al., 5 May 2025)

The information-oriented vehicular line of work is distinctive because it recasts ISCC resource management as a resource–information substitution problem. The Information-oriented Resource Trading Platform separates the information-production side from the information-sales side, measures efficiency through net profit $\Pi = R_{\text{info}} - C_{\text{TSFC}}$, and embeds the employment topology of IoAV into a graph neural network architecture. Two worker networks, distributor and purchaser, are trained in an asynchronous advantage GNN framework to balance information gain and TSFC resource consumption (Chen et al., 2024).

In learning-centric ISCC, alternating optimization is common. In Air-FEEL, the per-round design variables include the AirComp receive magnitude, local and global batch sizes, sensing power, and CPU frequency. The optimization is decomposed into a batch-size subproblem and a resource-allocation subproblem, both convex after reparameterization, and yields structural results such as $P_{k,s}^\star \propto 1/b_k$ and $f_k^\star$ equal to the minimum CPU frequency that makes the latency constraint tight (Wen et al., 21 Aug 2025). In distributed three-tier offloading, ADMM is used to optimize relaxed offloading decisions, while beamforming is handled through fractional programming, WMMSE reformulation, and SCA under sensing SINR constraints (Liu et al., 2024). In multidevice action-recognition ISCC, ADMM again separates device-side sensing-accuracy maximization from edge-side compute allocation, and the edge update admits a closed-form projection onto the CPU-budget constraint (Chen et al., 5 May 2025).

These examples show that there is no single canonical ISCC solver. The choice of method depends on whether the dominant difficulty is mixed-integer offloading, waveform–beamformer coupling, queueing and resource sharing, or learning-driven nonconvexity.

5. Application domains and neighboring extensions

ISCC has been applied to several 6G-motivated domains. In vehicular systems, the concept is tied to information acquisition for the Internet of Automatic Vehicles. Connected automatic vehicles use hybrid sensing modules that combine Active Wireless Sensing and Passive Wireless Sensing, offload data through mmWave links, and perform target-state extraction at local or edge units; ISCC coordinates these operations so that enough information is obtained for safe driving with minimal TSFC resource consumption (Chen et al., 2024). More broadly, autonomous driving is repeatedly cited as a core use case because it combines stringent latency, high reliability, dense sensing, and edge/cloud decision-making (Zhu et al., 2022, Xing et al., 2023).

Human motion recognition has become a common ISCC benchmark. One line uses FMCW radar-like sensing and ResNet-10 in Air-FEEL, with numerical results on a human motion recognition task showing how sensing SNR, AirComp distortion, and batch size jointly shape convergence and test accuracy (Wen et al., 21 Aug 2025). Another line studies wireless federated learning and split edge inference for human motion recognition using ISAC devices, local feature extraction, and task-oriented resource allocation across sensing, communication, and computation (Xing et al., 2023). A third line proposes an action-detection front end that offloads only time windows filled with signals of interest to the edge server, thereby reducing sensing-task overhead in multidevice ISCC (Chen et al., 5 May 2025). Across these works, human activity recognition functions as a tractable yet computation-intensive proxy for broader edge-intelligence workloads.

Industrial and robotic settings provide a different emphasis. The STAR-RIS-aided Internet of Robotic Things framework uses a full-duplex base station that simultaneously receives computation-offloading signals from decision robots and senses a target robot, while a STAR-RIS shapes the uplink channels; the corresponding objective is sum computation rate maximization under base-station power, sensing SNR, and STAR-RIS coefficient constraints (Li et al., 2024). Smart factories, networked robotics, and industrial control are also cited in survey work as environments where fast perception–decision loops require tightly integrated sensing, communication, and computation rather than conventional communication-only service (Zhu et al., 2022, Lee et al., 24 Jun 2025).

Low-altitude and aerial systems extend the ISCC idea toward closed-loop control. In the UAV-swarm literature, ISCCC is explicitly presented as the evolution from ISCC by adding control as a first-class component, yielding a closed-loop structure in which sensing acquires swarm and environmental state, communication transports state and control signals, computing performs estimation and planning, and control actuates physical motion (Ma et al., 8 Dec 2025). This suggests that, in some application domains, ISCC is best understood as the precursor to tighter cyber-physical integration rather than as an endpoint.

6. Robustness, security, limitations, and open directions

As ISCC systems migrate into safety- and mission-critical applications, the literature increasingly distinguishes between robustness and resilience. In this context, robustness denotes the ability to maintain performance under uncertainties such as fading, sensing noise, workload fluctuations, or model mismatch, whereas resilience denotes the ability to sustain a minimum service level under major disruptions such as node failures, outages, or attacks and then recover over time (Lee et al., 24 Jun 2025). This distinction matters because the design tools are different: robust ISCC emphasizes uncertainty-aware optimization, while resilient ISCC emphasizes redundancy, reconfiguration, coded transmission and computation, distributed multi-tier architectures, digital twins, and resilience-by-design (Lee et al., 24 Jun 2025).

Security has become a parallel concern. In low-altitude wireless networks, one recent formulation defines beampattern error, secrecy rate, and Age of Information as the sensing, secure-communication, and computing metrics of a security-aware ISCC problem, and optimizes BS sensing power, communication power, artificial-noise power, and computing rates through a DQN-based multi-objective evolutionary algorithm (Wang et al., 3 Nov 2025). The coupling is explicit: secure communication rate enters the AoI model as a service-rate parameter, so secrecy degradation directly worsens information freshness (Wang et al., 3 Nov 2025). This suggests that secure ISCC is not reducible to adding a cryptographic overlay; in many formulations it reshapes the sensing–communication–computing tradeoff itself.

Current models remain stylized. In ISCC-enabled Air-FEEL, the assumptions include a homogeneous target and homogeneous data distributions across devices, linear radar processing with Gaussian clutter and sensing noise, synchronous FEEL, fixed device scheduling, and gradient-descent-based training with a fixed-step schedule (Wen et al., 21 Aug 2025). In three-tier offloading formulations, common assumptions include perfect CSI, single-target sensing per terminal, binary offloading, static topology, and the absence of queueing or long-term arrival models (Liu et al., 2024). These assumptions are useful for tractable first-order analysis, but they limit direct extrapolation to asynchronous, mobile, heterogeneous, and imperfectly observed deployments.

Open problems therefore span both theory and implementation. Survey work points to the need for fundamental limits of task-oriented communication and ISCC, tighter convergence bounds for FEEL and Air-FEEL under AirComp and sensing errors, multimodal and dynamic sensing models, and cross-layer integration with RIS, THz, and massive MIMO (Zhu et al., 2022). Flexible-antenna research adds orientation-dependent channel modeling, near-field/far-field unified models, scalable joint optimization of boresight and beamforming, and learning-based control of spatially reconfigurable ISCC resources (Xiong et al., 16 Jun 2025). Robustness and resilience research adds self-organizing multi-tier architectures, holistic slicing over sensing, communication, and computation, digital-twin-assisted recovery, and resilience bottleneck identification as central agenda items for future ISCC networks (Lee et al., 24 Jun 2025).

Taken together, the literature portrays ISCC as a task-oriented systems framework for 6G in which sensing, communication, and computation are jointly modeled, jointly optimized, and increasingly jointly secured. Its core technical signature is not any single waveform or algorithm, but the insistence that sensing quality, communication behavior, and computation placement or complexity must be analyzed as parts of one end-to-end pipeline.