Integrated Sensing, Communication & Computation

• ISCC is a 6G-oriented framework that tightly integrates sensing, communication, and computation to meet end-task objectives such as latency, accuracy, and energy efficiency.
• It co-designs waveform, resource management, and AI-driven processing to jointly optimize sensing fidelity, communication rate, and computation latency.
• ISCC architectures leverage edge, cloud, and emerging technologies like RIS, UAVs, and fluid antennas to enable real-time, task-driven decision making.

Searching arXiv for recent ISCC papers to ground the article in current literature. arxiv_search.query({"^^^^^^^^12search_query12 Sensing, Communication, and Computation\"^^^^^^^^12 OR all:ISCC12^^^^^^^^", "^^^^^^^^12start12 ^^^^12search_query12^^^^, "^^^^^^^^12max_results12 ^^^^12all:\12search_query12^^^^, "^^^^^^^^12sort_by12 "^^^^^^^^12submittedDate12 "^^^^^^^^12sort_order12 "^^^^^^^^12descending12 arxiv_search.query({"^^^^^^^^12search_query12 Sensing, Communication, and Computation\"^^^^^^^^12all:\12all:\12 Sensing, Communication, and Computation\"", "^^^^^^^^12start12 ^^^^12search_query12^^^^, "^^^^^^^^12max_results12 ^^^^12all:\12search_query12^^^^, "^^^^^^^^12sort_by12 "^^^^^^^^12submittedDate12 "^^^^^^^^12sort_order12 "^^^^^^^^12descending12 Integrated Sensing, Communication, and Computation (ISCC) is a ^^^^12submittedDate12^^^^G-oriented, task-oriented design paradigm that co-designs sensing for information acquisition, communication for information exchange, and computation for information processing and decision making as a single coupled system. In ISCC, raw or partially processed sensing data are not treated as the endpoint of the wireless pipeline; instead, sensing outputs are uploaded, fused, and computed at devices, edge servers, or cloud resources to meet end-task objectives such as inference accuracy, latency, energy efficiency, and reliability. Relative to pairwise integrations such as integrated communication and computation (ICC), integrated sensing and computation (ISC), and integrated sensing and communication (ISAC), ISCC explicitly closes the loop across all three modules and optimizes shared radio, hardware, and compute resources against task-level utility rather than isolated module-level surrogates (&&&^^^^12search_query12^^^^&&&, &&&^^^^12all:\12^^^^&&&, &&&^^^^12 OR all:ISCC12^^^^&&&).

^^^^12all:\12^^^^. Conceptual foundations and relation to adjacent paradigms

ISCC is commonly defined as a holistic integration of sensing, communication, and computation pipelines such that the quality-of-service of each module is tuned to the end task and resources are co-managed across modules and devices (&&&^^^^12all:\12^^^^&&&). In this formulation, sensing provides priors for communication and data for computing; communication fuses multi-agent sensing and coordinates distributed computing; computation enhances sensing through AI-aided processing and can also reduce communication burden through semantic or feature transmission (&&&^^^^12search_query12^^^^&&&). This coupling is especially consequential in latency-sensitive sensing tasks, because delayed inference can negate the value of sensed data for safety-critical decisions (&&&^^^^12 OR all:ISCC12^^^^&&&).

The paradigm is routinely positioned as a generalization of ICC, ISC, and ISAC. ICC covers task offloading, MEC, DAG-based scheduling, and over-the-air computation; ISC covers radar and wireless sensing, multi-modal sensing, and mobile crowdsensing; ISAC covers radar-communications coexistence and dual-functional radar-communications waveform design. ISCC subsumes these by enabling full resource sharing across the three modules, aligning module-level design with task-level utility, and explicitly optimizing tight cross-domain couplings such as the effect of sensing fidelity on feature quality, of communication rate on deliverable features or updates, and of compute speed on end-to-end latency (&&&^^^^12search_query12^^^^&&&). In the ^^^^12submittedDate12^^^^G edge-intelligence literature, this is the central distinction between task-agnostic wireless design and task-oriented ISCC (&&&^^^^12sort_order12^^^^&&&).

A recurring theme in the literature is that ISCC should be evaluated by end-task performance rather than by throughput, sensing resolution, or compute speed in isolation. For edge AI, separate optimization can create objective inconsistency: maximizing throughput does not necessarily maximize inference accuracy, and maximizing sensing resolution may increase payload size and compute load beyond latency budgets. ISCC reframes the system objective around joint accuracy-latency-energy trade-offs for learning, inference, control, and perception tasks (&&&^^^^12descending12^^^^&&&, &&&^^^^12all:\12^^^^&&&).

^^^^12 OR all:ISCC12^^^^. Architectures, entities, and end-to-end pipeline

Canonical ISCC architectures include base stations or access points co-located with edge servers, mobile devices with sensing modalities such as radar, camera, LiDAR, microphones, or GPS, and optional auxiliary infrastructure such as reconfigurable intelligent surfaces (RIS), UAV relays, roadside units, satellites, and cloud servers (&&&^^^^12search_query12^^^^&&&). Several papers also study explicit three-tier cloud-edge-terminal architectures, where terminals can process tasks locally, offload to MEC servers, or further offload through the edge to a cloud server depending on latency and compute availability (&&&^^^^12all:\12all:\12^^^^&&&, &&&^^^^12all:\12 OR all:ISCC12^^^^&&&).

The standard end-to-end pipeline is

PRESERVED_PLACEHOLDER_^^^^12search_query12^^^^

with optional local preprocessing or feature extraction before transmission and optional feedback or actuation after edge/cloud computation (&&&^^^^12search_query12^^^^&&&). In practical systems this pipeline may be instantiated as sensing PRESERVED_PLACEHOLDER_^^^^12all:\12^^^^ local preprocessing or feature extraction PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12^^^^ uplink communication or AirComp aggregation PRESERVED_PLACEHOLDER_^^^^12start12^^^^ edge/cloud inference or control PRESERVED_PLACEHOLDER_^^^^12max_results12^^^^ downlink feedback (&&&^^^^12search_query12^^^^&&&, &&&^^^^12 OR all:ISCC12^^^^&&&). In edge-AI settings, the communicated object may be raw samples, spectrograms, model updates, intermediate features, or semantic representations rather than conventional bit streams (&&&^^^^12all:\12^^^^&&&, &&&^^^^12descending12^^^^&&&).

Communication and offloading are usually modeled with standard Shannon-type or MIMO rates. A representative expression is

PRESERVED_PLACEHOLDER_^^^^12sort_by12^^^^

while computation latency is typically represented as PRESERVED_PLACEHOLDER_^^^^12submittedDate12^^^^ and dynamic compute energy as PRESERVED_PLACEHOLDER_^^^^12sort_order12^^^^ (&&&^^^^12search_query12^^^^&&&). In MEC-centric ISCC, local and edge execution are often contrasted through formulas such as

PRESERVED_PLACEHOLDER_^^^^12descending12^^^^

with transmission time PRESERVED_PLACEHOLDER_^^^^12search_query12^^^^ (&&&^^^^12 OR all:ISCC12^^^^&&&).

The architecture can be centralized, federated, or hybrid. Centralized edge learning uploads sensed data to an edge server for training; federated edge learning keeps raw data local but exchanges model updates; split inference or split learning partitions the model between the device and the edge; and AirComp-based realizations collapse communication and aggregation into a single analog superposition step (&&&^^^^12sort_order12^^^^&&&, &&&^^^^12all:\12^^^^&&&). This suggests that ISCC is not a single protocol stack but a system-theoretic umbrella covering multiple deployment styles so long as sensing, communication, and computation are jointly designed against a shared task objective.

^^^^12start12^^^^. Canonical models, metrics, and optimization structure

ISCC optimization typically combines communication metrics, sensing metrics, and computation or latency metrics in a single problem. A representative multi-objective form is

PRESERVED_PLACEHOLDER_^^^^12all:\12search_query12^^^^

where PRESERVED_PLACEHOLDER_^^^^12all:\12all:\12^^^^ may include power, beamforming, bandwidth, time allocation, CPU frequencies, offloading fractions, or other control variables; PRESERVED_PLACEHOLDER_^^^^12all:\12 OR all:ISCC12^^^^ may denote Fisher information, negative CRB, beampattern gain, INR, or detection probability; and PRESERVED_PLACEHOLDER_^^^^12all:\12start12^^^^ may denote module-level or end-to-end latency (&&&^^^^12search_query12^^^^&&&). RIS-assisted formulations use the same structure with explicit dependence on beamformers PRESERVED_PLACEHOLDER_^^^^12all:\12max_results12^^^^, RIS phases PRESERVED_PLACEHOLDER_^^^^12all:\12sort_by12^^^^, and offloading or scheduling variables PRESERVED_PLACEHOLDER_^^^^12all:\12submittedDate12^^^^, together with constraints such as PRESERVED_PLACEHOLDER_^^^^12all:\12sort_order12^^^^, PRESERVED_PLACEHOLDER_^^^^12all:\12descending12^^^^, PRESERVED_PLACEHOLDER_^^^^12all:\12search_query12^^^^, and PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12search_query12^^^^ (&&&^^^^12 OR all:ISCC12^^^^&&&).

On the sensing side, the literature repeatedly uses CRB, detection probability, false-alarm probability, range resolution, Doppler resolution, and beampattern metrics. Standard expressions include

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12all:\12^^^^

and, under Gaussian hypothesis testing,

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12 OR all:ISCC12^^^^

as well as classical range and Doppler resolutions such as PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12start12^^^^ and PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12max_results12^^^^ (&&&^^^^12search_query12^^^^&&&, &&&^^^^12 OR all:ISCC12^^^^&&&, &&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&). For radar-style monostatic sensing, the received power is often written as

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12sort_by12^^^^

with modifications when RIS or orientation control strengthens the effective echo path (&&&^^^^12 OR all:ISCC12^^^^&&&).

On the communication-computation side, AirComp introduces explicit computation-quality metrics. A representative model is

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12submittedDate12^^^^

with aggregation error

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12sort_order12^^^^

optimized by joint transmit and receive beamforming or scaling (&&&^^^^12search_query12^^^^&&&). In OFDM AirComp-based ISCC, the objective can be to minimize computational MSE by jointly designing the transmitting vector and the aggregation vector under sensing constraints (&&&^^^^12 OR all:ISCC12search_query12^^^^&&&). In beamforming-and-offloading formulations, weighted minimum mean-square error reformulations and fractional programming are used to convert latency-minimization objectives into tractable alternating updates (&&&^^^^12all:\12all:\12^^^^&&&, &&&^^^^12all:\12 OR all:ISCC12^^^^&&&).

Solution methods follow the nonconvex structure of these problems. Deterministic approaches include semidefinite relaxation, successive convex approximation, majorization-minimization, ADMM, block coordinate descent, primal-dual methods, and WMMSE reformulations (&&&^^^^12 OR all:ISCC12^^^^&&&, &&&^^^^12search_query12^^^^&&&, &&&^^^^12all:\12all:\12^^^^&&&, &&&^^^^12all:\12 OR all:ISCC12^^^^&&&). Learning-based control appears when channels, tasks, or mobility are highly dynamic. In RIS-assisted ISCC, for example, deep reinforcement learning is used to map state variables such as task sizes, channel information, and user locations to actions including offloading decisions, RIS phase shifts, and beamforming vectors with reward defined by negative user energy consumption under latency and QoS constraints (&&&^^^^12 OR all:ISCC12^^^^&&&). Similar RL structures appear in information-centric IoAV resource management and dynamic access control for orientation-aware systems (&&&^^^^12all:\12all:\12^^^^&&&, &&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&).

^^^^12max_results12^^^^. Task-oriented ISCC for edge learning and inference

A major strand of ISCC research treats sensing, communication, and computation as a single edge-intelligence pipeline. In edge AI, devices first sense task-relevant data, then either train local models, send updates, offload features, or perform split inference. The central claim is that the performance of edge learning and edge inference depends jointly on sensing quality, communication rate or distortion, and computation latency, so these modules must be scheduled together (&&&^^^^12all:\12^^^^&&&). This view is developed in both application-layer and physical-layer terms, including digital and analog FEEL, dual-functional and triple-functional beamforming, and semantic or feature-oriented transmission (&&&^^^^12descending12^^^^&&&).

For federated learning and edge learning, ISCC introduces sensing directly into the convergence analysis. In over-the-air federated edge learning, devices sense objects, compute local gradients, and aggregate them in one shot via AirComp. The convergence bound explicitly contains sensing noise, residual clutter, AirComp distortion, batch size, and computation latency terms, showing that sensing, communication, and computation compete for the same latency and energy budgets (&&&^^^^12max_results12all:\12^^^^&&&). A key design outcome is that less sensing power should be consumed if a larger batch of data samples is obtained, while the optimal computation speed is the minimum one that satisfies the latency constraint for a given batch size (&&&^^^^12max_results12all:\12^^^^&&&).

For edge inference, split computing and feature transmission are central. One line of work models inference accuracy as a function of sensing SNR, communication distortion, and compute frequency, then jointly optimizes offloading, bandwidth, power, and CPU allocation under delay and energy constraints (&&&^^^^12all:\12^^^^&&&). Another line introduces adjustable split inference, model pruning, and feature quantization, and derives an explicit approximation of inference accuracy from pruning-induced feature extraction error and quantization error. The resulting design can reduce energy consumption by up to ^^^^12max_results12search_query12^^^^% compared to existing methods, particularly in low-latency scenarios (&&&^^^^12max_results12max_results12^^^^&&&). These formulations make the task-oriented aspect of ISCC concrete: split layer, pruning ratio, quantization bits, sensing power, transmit power, and local CPU frequency all become first-class optimization variables.

Recent work also generalizes ISCC beyond single-modality sensing. In task-oriented multimodal edge intelligence, each device extracts compact, discriminative features from its modality and transmits them to the edge for joint inference. The maximal coding rate reduction criterion is used both to train device-side extractors and to define an edge-side sensing metric through covariance-aware log-determinant expressions that capture inter-class compactness and inter-modal correlation. On a human activity recognition task with WiFi, RFID, and mmWave radar, this multimodal ISCC framework outperforms device-level quantization, average time allocation, and single-modality baselines under limited resource conditions (&&&^^^^12max_results12sort_by12^^^^&&&). A plausible implication is that ISCC increasingly functions as a unifying framework for task-aware feature formation, transmission, and fusion rather than only for waveform sharing.

UAV-enabled split federated learning extends the same logic to airborne sensing platforms. There the model is partitioned across UAVs and edge servers, client-side aggregation occurs less frequently than server-side aggregation, and convergence bounds depend on the split point, sensing volume, client aggregation interval, and UAV deployment through sensing success probabilities and air-to-ground rates (&&&^^^^12max_results12submittedDate12^^^^&&&). This closes the loop between physical placement, sensing quality, communication efficiency, and distributed training dynamics.

^^^^12sort_by12^^^^. Enabling technologies, reconfigurable architectures, and deployment scenarios

RIS is one of the most extensively studied physical enablers for ISCC. In RIS-assisted ISCC, BSs or APs with MEC servers perform communication and radar sensing while RIS panels with dozens to hundreds of low-cost passive reflecting elements reshape the propagation environment. The phase-shift matrix is commonly written as

PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12descending12^^^^

with ideal passive constraints PRESERVED_PLACEHOLDER_^^^^12 OR all:ISCC12search_query12^^^^ or discrete-phase constraints under quantization (&&&^^^^12 OR all:ISCC12^^^^&&&). RIS creates virtual LoS paths, strengthens blocked links, improves radar echo SNR and beampattern focus, increases communication rate for faster offloading, and therefore reduces MEC latency and energy (&&&^^^^12 OR all:ISCC12^^^^&&&). Case studies with a BS using ^^^^12all:\12submittedDate12^^^^ antennas, a ^^^^12max_results12search_query12^^^^-element RIS, and ^^^^12all:\12submittedDate12^^^^ users show larger desired-direction beampattern gain and lower user energy consumption than no-RIS baselines, with benefits increasing as the number of RIS elements grows (&&&^^^^12 OR all:ISCC12^^^^&&&).

Non-fixed antenna architectures extend reconfigurability from the environment to the radio front end. Intelligent rotatable antennas provide a spatially reconfigurable radiation pattern by rotating the antenna boresight in three dimensions through mechanical, electronic, or hybrid actuation. In ISCC they affect communication rate, sensing SNR and resolution, and offloading latency simultaneously through orientation-dependent gains PRESERVED_PLACEHOLDER_^^^^12start12search_query12^^^^ and PRESERVED_PLACEHOLDER_^^^^12start12all:\12^^^^ (&&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&). Fluid antennas, by contrast, move the receive element position within a bounded aperture and thereby exploit spatial sampling diversity and interference suppression. In vehicular ISCC, joint optimization of fluid-antenna positions, receive combining, and MEC CPU allocation reduces end-to-end latency across sensing, communication, and computation relative to fixed-antenna baselines (&&&^^^^12sort_by12all:\12^^^^&&&). These technologies are complementary rather than mutually exclusive: the literature also identifies hybrid IRA-RIS systems as an open direction (&&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&).

Deployment scenarios concentrate on high-mobility and coverage-limited environments. UAV-assisted ISCC leverages strong LoS channels and flexible placement; UAVs can carry sensing and edge-compute payloads and act as airborne MEC nodes for emergency communications or precision agriculture (&&&^^^^12 OR all:ISCC12^^^^&&&). Internet of Vehicles and IoAV scenarios emphasize rapidly varying channels, mmWave or THz blockage, roadside RIS deployment, and joint planning of sensing and MEC offloading for safe autonomous driving (&&&^^^^12 OR all:ISCC12^^^^&&&, &&&^^^^12all:\12all:\12^^^^&&&). In information-centric IoAV, time-space-frequency-computing resources are decoupled into universal resource pools and managed through an Information-oriented Resource Trading Platform that optimizes information gain against TSFC resource cost, with topology-aware asynchronous advantage graph neural networks used for dynamic allocation (&&&^^^^12all:\12all:\12^^^^&&&).

Broader ^^^^12submittedDate12^^^^G integration places ISCC in digital twins, computing power networks, and space-air-ground integrated networks. Digital twins use federated learning, federated split learning, and centralized learning across heterogeneous sensing and compute resources; computing power networks share microservices such as sampling, filtering, and feature extraction across tasks; and SAGIN variants combine terrestrial, aerial, and satellite components for wide-area sensing and remote computation (&&&^^^^12search_query12^^^^&&&). This suggests that ISCC is increasingly treated not only as a radio design problem but also as a system orchestration problem spanning device, edge, cloud, and sometimes orbital layers.

^^^^12submittedDate12^^^^. Trade-offs, robustness, and open research problems

The principal difficulty of ISCC is not the existence of individual sensing, communication, or computation mechanisms, but the fact that the three functions compete for time, energy, bandwidth, spatial degrees of freedom, and compute cycles. Shared waveform design must satisfy both radar and communication requirements; offloading and MEC scheduling must meet sensing freshness constraints; overloaded servers and congested channels can degrade both latency and sensing accuracy; and higher sensing fidelity can enlarge feature payloads and inference cost (&&&^^^^12 OR all:ISCC12^^^^&&&, &&&^^^^12search_query12^^^^&&&, &&&^^^^12descending12^^^^&&&). The resulting design space is naturally Pareto-structured rather than single-objective.

Robustness and resilience formalize these concerns under uncertainty and disruption. Robustness refers to maintaining performance under routine uncertainties, whereas resilience denotes sustaining a minimum level of service under major disruptions and then recovering (&&&^^^^12submittedDate12all:\12^^^^&&&). In ISCC this distinction is sharpened by tight cross-module dependencies: channel fading, estimation errors, workload variation, topology changes, jamming, or blockages can simultaneously affect detection, communication reliability, and compute timeliness. Relevant metrics include outage probability, service reliability, detection and false-alarm behavior, resistance during disruptions, worst-case service dips, detection-reaction latency, time-to-recover, and mean time to failure (&&&^^^^12submittedDate12all:\12^^^^&&&). A distributed radar case study with target-location uncertainty bounded by PRESERVED_PLACEHOLDER_^^^^12start12 OR all:ISCC12^^^^ m and line-of-sight blockages shows that combining worst-case provisioning with obstruction-aware reconfiguration maintains the normalized sensing SNR above threshold across the entire trajectory, whereas robustness-only or resilience-only strategies fail in specific segments (&&&^^^^12submittedDate12all:\12^^^^&&&).

Open problems recur across the literature. They include unified task-oriented information theory for joint sensing-communication-computation, scalable online optimization under large heterogeneous networks, robust modeling across near-field sensing and far-field offloading regimes, service continuity under fast mobility, privacy and secure control signaling, standardization of APIs and protocols, and field experimentation with large-scale benchmarks (&&&^^^^12search_query12^^^^&&&, &&&^^^^12submittedDate12all:\12^^^^&&&, &&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&). Hardware-specific issues include RIS phase quantization, beam squint in wideband operation, RF nonidealities, servo and MEMS reliability for rotatable antennas, and actuation energy for movable or fluid antennas (&&&^^^^12 OR all:ISCC12^^^^&&&, &&&^^^^12 OR all:ISCC12submittedDate12^^^^&&&, &&&^^^^12sort_by12all:\12^^^^&&&). At the application layer, multimodal sensing, continual adaptation, and joint sensing-communication-computation semantics remain underdeveloped (&&&^^^^12max_results12sort_by12^^^^&&&, &&&^^^^12sort_order12^^^^&&&).

Across these strands, a stable conclusion emerges: ISCC is best understood as a cross-layer, task-oriented framework in which sensing, communication, and computation are not merely co-located but mathematically and operationally entangled. Its mature formulations therefore combine physical-layer waveform or beam design, communication and offloading control, compute scheduling, and task-level evaluation in a single optimization or learning loop (&&&^^^^12search_query12^^^^&&&, &&&^^^^12 OR all:ISCC12^^^^&&&).