AI-Aided and AI-Oriented Integrated Sensing, Communication, and Computation: A Survey
Organizations: Department of Computer Science, Université du Québec à Montréal (UQAM), Montréal, Canada · College of Computing, Mohammed VI Polytechnic University, Benguerir, Morocco
Abstract
The evolution from 5G to 6G networks marks a transition toward the "intelligence of everything." While 5G enables enhanced connectivity, 6G aspires to embed intelligence directly into communication infrastructures, creating systems that are adaptive, context aware, and capable of autonomous decision making. Realizing such systems depends on harmonizing three fundamental functions that compete for shared network resources and often pursue conflicting goals: (i) sensing for environmental and contextual information acquisition, (ii) communication for reliable and timely information exchange, and (iii) computation for information processing and decision making. This harmonization necessitates shifting from independent optimization to their joint design as a unified system, a paradigm known as Integrated Sensing, Communication, and Computation (ISCC). In parallel, AI has emerged both as an enabling tool for ISCC optimization and as a workload whose requirements influence ISCC system design. This survey comprehensively examines ISCC from this dual perspective by distinguishing between AI-aided ISCC, in which learning-based methods support the joint design and optimization of sensing, communication, and computation, and AI-oriented ISCC, in which these functions are jointly configured to support learning, inference, and distributed intelligence. The survey further examines the co-design of AI and ISCC, where AI requirements and ISCC constraints jointly shape system and model decisions. It further provides a comparative synthesis of the resulting design approaches, system architectures, performance metrics, and evaluation practices, relates these developments to application-specific requirements across different domains, and identifies open research problems toward scalable, adaptive, and trustworthy AI-ISCC systems.
Figures & tables
| Ref. | Year | Primary Focus | Coverage | AI-Aided | AI-Oriented | Dual Perspective | AI–ISCC Co-design | Performance Metrics | System Architectures | Application Domains |
| D. Wen et al. [ 43 ] | 2025 | Comprehensive ISCC fundamentals and enabling technologies | - Analysis of partial integration (ICC, ISC, ISAC) with limitations - Signal design and waveform optimization for joint integration - Network resource management strategies - 6G enabling technologies and applications | ✗ | ✗ | ✗ | ✗ | |||
| G. Zhu et al. [ 44 ] | 2023 | ISCC paradigm for enabling edge intelligence and AIoT in 6G | - Task-oriented ISCC design principles for edge AI - Federated edge learning system architectures - Edge AI inference schemes (on-device, on-edge, split) - Design challenges and enabling techniques | ✗ | ✓ | ✗ | ✓ | ✓ | ||
| G. Wu et al. [ 45 ] | 2025 | Overview of ISCC with AI integration | - Review of sensing, communication, and computation technologies and partial integrations - Analysis of the interactions between AI and ISCC - Intelligent resource management approaches - ISCC support for AI applications | |||||||
| Y. Zhang et al. [ 46 ] | 2026 | AI-empowered task-oriented sensing, communication, and computation | - AI-based and AI-assisted methods for task-oriented SCC - Task classification by modality, scenario, and task attributes - Model-driven and task-execution performance metrics - Applications and technical challenges of AI-empowered TSCC | ✓ | ✓ | ✓ | ||||
| Y. Liu et al. [ 47 ] | 2026 | Integrated sensing and edge AI for intelligent perception | - Integration of sensing, communication, and AI computation - Edge AI architectures and task-oriented sensing - Performance metrics and system tradeoffs - Applications, enabling technologies, and standardization | ✓ | ✓ | ✓ | ✓ | |||
| C. Li et al. [ 48 ] | 2026 | ISCC for Internet of Vehicles and connected autonomous driving | - Three-phase ISCC evolution (Collaborative, Fusion, Integrated) - Physical-layer signal integration techniques - Network-level task-oriented resource management - Safety-critical and real-time requirements for vehicular systems | ✗ | ✗ | ✗ | ✓ | ✓ |
| Ref. | Scenario | Method | Decision Variables | Objective / Main Constraints | Results |
| Joint resource allocation and management | |||||
| [ 71 ] | Collaborative sensing in IoV | DRL | Uplink-bandwidth allocation, downlink-bandwidth allocation, and edge-computing allocation | Maximize collaborative sensing-task completion rate subject to service-delay and resource constraints | Higher sensing-task completion rate than the conventional allocation schemes considered in the paper |
| [ 70 ] | Heterogeneous vehicular intelligent tasks | DDQN | Communication-resource, computation-resource, and cache allocation for real-time and delay-tolerant tasks | Optimize task performance under heterogeneous resource and latency requirements | At least about higher exploration efficiency and approximately higher mAP under resource constraints |
| [ 72 ] | Single-UAV ISCC | D3PG-Light | Communication-bandwidth fraction , sensing-time fraction , and CPU-frequency fraction | Reduce latency while maintaining sensing quality and controlling energy consumption | th-percentile latency reduced from ms to – ms; fewer than k parameters and fewer parameters than original D3PG |
| [ 73 ] | UAV-assisted emergency IoV | ACBC-TD3 | Vehicle scheduling order, UAV motion, and task-offloading ratio | Maximize collaborative sensing-task completion under latency and emergency-link constraints | Faster convergence and higher task-completion performance than standard TD3 under stochastic emergency conditions |
| Waveform and beamforming design | |||||
| Ref. | AI Workload | Sensing Variables | Communication Variables | Computation / Model Variables | AI Objective / Metric | Results |
| Federated edge learning | ||||||
| [ 63 ] | Over-the-air FEEL | Per-device sensing power and acquired sample size | Transmit power and AirComp receive magnitude | CPU frequency and local/global batch sizes | Per-round loss degradation and convergence | Larger batches permit lower sensing power; for a fixed batch, the optimal CPU frequency is the minimum satisfying the latency constraint. Human-motion experiments verify the convergence analysis and gains over fixed-resource designs. |
| [ 81 ] | UAV-assisted FEEL | UAV deployment and sensing batch size | Per-UAV bandwidth allocation | Local training workload determined by batch size | Training time, convergence, and testing accuracy | The proposed bandwidth, batch-size, and position optimization improves convergence and testing accuracy over the evaluated baseline deployment/resource-allocation schemes. |
| [ 84 ] | FedAVG-ISCC / FedSGD-ISCC | Per-round sample collection and sensing errors | Wireless update exchange, communication errors, and aggregation parameters | Local update count, CPU resources, and learning protocol | Convergence and test accuracy under IID/non-IID data | FedAVG-ISCC performs better under IID data, whereas FedSGD-ISCC is more robust to non-IID data and communication errors; the ISCC variants also reduce latency and energy relative to the OTA-FL counterparts considered. |
| [ 86 ] | Online FEEL | fresh/historical data composition | OFDM resources for model aggregation | Cache allocation, local computation, and cached-data selection | Per-round loss reduction and convergence | Loss-aware caching outperforms the evaluated cache and resource-allocation baselines while accounting for concept drift in historical samples. |
| Edge, split, and collaborative inference | ||||||
| Ref. | AI-Side Requirement or Adaptation | ISCC Decisions | Coupling | Results |
| AI requirements driving ISCC design | ||||
| [ 93 ] | FEEL convergence under IID and non-IID data | Transmit beamforming and flexible sensing/communication time allocation | The convergence analysis determines the resource objective; sensing and communication resources control the data and model-update process that determines subsequent convergence | Improved model-training performance under limited resources for both IID and non-IID data relative to the evaluated alternatives |
| [ 94 ] | Learning-performance gain of multimodal federated perception | Time, frequency, and computation-resource scheduling | The marginal learning benefit of resource expansion is balanced against sensing, model-transmission, and training costs | Experimental evaluation demonstrates improved learning/resource efficiency and robustness of the proposed service-oriented scheduling mechanisms |
| ISCC constraints driving AI adaptation | ||||
| [ 7 ] | DNN split point for UAV–edge target classification | User association, target assignment, beamforming, computation allocation, and UAV locations | The split determines local/edge workloads and intermediate transmission, while latency, accuracy, radio, and computation conditions determine the feasible split | Adaptive splitting outperforms fixed collaborative-computation configurations in the reported simulations |
| [ 96 ] | On-device, on-server, or cooperative inference mode | Mode-dependent sensing, communication, and computation resources | Each inference mode induces different communication and computation costs; the resource state determines the energy-efficient mode under an accuracy guarantee | Practical-platform experiments validate the theoretical mode-selection behavior across operating conditions |
| Architecture | Sensing Locus | Information Flow | Computation Locus | Representative Works |
|---|---|---|---|---|
| Centralized device–edge | Sensing is performed at distributed edge devices | Devices transmit features, embeddings, or model updates to a common edge server; AirComp may aggregate concurrent transmissions | Feature extraction or local training at devices; aggregation, inference, or global-model update at the edge | [ 87 , 103 , 63 , 91 ] |
| Hierarchical device–edge–cloud | Observations originate at sensing devices | Intermediate DNN representations are transferred from device to MEC and, when selected, from MEC to cloud | Different DNN segments are executed at device, MEC, and cloud tiers | [ 97 ] |
| Decentralized D2D | Each participating device senses and extracts local features | Intermediate features are exchanged directly through D2D links and aggregated among cooperating devices | Feature extraction, aggregation, and inference are distributed across peer devices | [ 89 ] |
| Aerial and mobile | Ground devices, targets, and/or aerial platforms participate in sensing | Information is exchanged over ground–air, air–ground, or air–edge links whose geometry changes with platform motion | Processing may be divided among ground devices, UAV/AAV platforms, and edge infrastructure | [ 81 , 90 , 104 ] |
| Cell-free distributed | Terminals perform sensing while geographically distributed access points provide wireless connectivity | Tasks are transmitted through distributed access points toward selected edge servers or the cloud | Execution may occur locally, at distributed edge servers, or in the cloud | [ 105 ] |
| Ref. | Physical / Resource Metrics | Representation Metrics | Learning / Inference Metrics | End-to-End / System Metrics |
|---|---|---|---|---|
| [ 103 ] | Sensing quality and wireless aggregation error | Minimum pairwise discriminant gain | Classification accuracy | Task-oriented inference performance |
| [ 39 ] | Sensing energy, communication energy, computation energy, transmission latency, and computation latency | Feature-quantization distortion and pruning-induced model degradation | Inference accuracy | Total device energy and end-to-end inference latency |
| [ 81 ] | Sensing success probability, communication rate, and training-time components | – | Training loss, convergence behavior, and testing accuracy | Federated-training completion time |
| [ 63 ] | Sensing distortion, aggregation error, per-round latency, and energy consumption | AirComp aggregation distortion | Training loss, convergence behavior, and testing accuracy | Learning performance under latency and energy budgets |
| [ 91 ] | Sensing distortion, communication latency, computation latency, and energy consumption | Embedding-aggregation distortion | Training loss and testing accuracy | Training performance under latency and energy constraints |
| [ 92 ] | Transmission latency and communication energy across sensing modalities | Maximal coding rate reduction and representation discriminability | Recognition accuracy | Task performance under communication-resource constraints |
| Dataset / Resource | Source | Signals / Modalities | Tasks / Labels | System Information | Scope / Representative Use |
|---|---|---|---|---|---|
| XRF55 [ 106 ] | Measured | Wi-Fi, RFID, mmWave, and Kinect observations | Human indoor activity recognition | Synchronized multimodal observations across sensing technologies | Used in full multimodal ISCC [ 92 ] |
| DeepSense 6G [ 107 ] | Measured | mmWave communication, radar, LiDAR, camera, GPS, and other scenario-dependent modalities | Beam prediction, positioning, blockage and environment-aware wireless tasks | Wireless measurements, position, mobility, and synchronized multimodal observations | Sensing–communication and ISAC research |
| DeepMIMO [ 108 ] | Ray tracing | Parameterized MIMO channel responses and multipath information | Beam prediction, localization, channel learning, and spatial inference | Tx/Rx positions, delays, path gains, AoA/AoD, and propagation geometry | Wireless AI and ISAC-relevant benchmarking |
| DISC [ 109 ] | Measured | 60-GHz IEEE 802.11ay channel impulse responses | Activity recognition, gait analysis, subject localization/tracking, and sparse sensing | Communication-native CIRs collected under uniform and traffic-like packet patterns | Direct communication-native ISAC |
| M 3 SC [ 112 ] | Synthetic | RGB, depth, LiDAR, mmWave radar waveforms/point clouds, and communication CIRs | Multimodal sensing and communication tasks | Aligned physical-scene and electromagnetic-domain data under varying scenarios | Multimodal sensing–communication integration |
| SynthSoM [ 113 ] | Synthetic | RF channels, path loss, mmWave radar, RGB, depth, and LiDAR | Multimodal perception, channel learning, and air–ground cooperative tasks | Aligned sensing and communication data across weather, mobility, frequency, and scenario variations | Multimodal sensing–communication integration |
Integrated Sensing and Communication
Integrated sensing and communication combines wireless data links with radar-like perception using shared signals, spectrum, and hardware.
See all 55 papers