Integrated Communication and Control (ICAC)

• Integrated Communication and Control (ICAC) is a multidisciplinary framework that jointly designs communication and control strategies to optimize system performance in cyber-physical environments.
• It employs mathematical models, algorithms, and optimization methods to co-optimize control objectives and communication quality, thus enhancing energy efficiency and reliability.
• Applications such as UAV swarms, industrial automation, and multi-agent reinforcement learning benefit from the integrated design, achieving significant performance gains and cost reductions.

Integrated Communication and Control (ICAC) encompasses system architectures, mathematical frameworks, algorithms, and optimization methods that intertwine the design and operation of communication subsystems with the requirements and dynamics of feedback control and estimation. ICAC is foundational for modern cyber-physical systems—including industrial automation, UAV swarms, and multi-agent reinforcement learning—where resource constraints, stochasticity, or the need for ultra-reliable low-latency operation preclude the classical separation of communication and control functions.

1. Foundational Concepts and Structural Models

ICAC systems capture the bidirectional and often tightly coupled relationships among actuation, sensing, computation, and message exchange. Canonical models integrate plant or agent dynamics, discrete or continuous time, and state-feedback or output-feedback control, with stochastic communication processes (delay, loss, quantization, fading) and real or virtual communication networking.

• Vector-State LQG Model: The state $s_k \in \mathbb{R}^r$ evolves via $s_{k+1} = F s_k + G u_k + w_k$; observation $y_k = H s_k + J u_k + v_k$. Both $w_k$ and $v_k$ are independent Gaussian noises; neither the controller nor observer accesses $s_k$ directly. The control input $u_k$ carries both corrective actions and, potentially, embedded messages for information transfer—a core ICAC capability (Jahangiri et al., 20 Dec 2025).
• Closed-Loop Sensing–Communication–Control Models: Discrete-time plant/actuator states are periodically sensed, quantized, and transmitted to a remote controller. This is reflected in equations such as $X_{t+1} = \tilde{A} X_t + \tilde{B} U_t$, with finely modeled uplink and downlink queues, delay, and quality-of-service (QoS), intimately coupling all resource and convergence constraints (Meng et al., 2024).
• Multi-Agent and Swarm Architectures: For UAV swarms, the agent trajectories and wireless communication (OFDMA resource assignment, channel selection) are jointly modeled as Markov Decision Processes (MDPs), capturing not just dynamics and control, but collision avoidance, RF link assignment, and real-time co-design (Sun et al., 28 Sep 2025, Wei et al., 11 Feb 2025).

2. Joint Optimization Objectives and Performance Metrics

A distinguishing feature of ICAC is co-optimizing control-system objectives and communication-related costs under a unified criterion.

• Energy Efficiency×Fairness: In UAV-enabled OFDMA networks with swarming, the per-slot efficiency is $\eta^{ee}(t) = \hat{f}(t) R_c(t)/E(t)$, where $E(t)$ is UAV swarm propulsion energy and $\hat{f}(t)$ is Jain's fairness index over cumulative communication rates. The joint objective is maximizing the long-term sum $\sum_t \eta^{ee}(t)$ subject to safety, assignment, and power constraints (Sun et al., 28 Sep 2025).
• LQG Cost under Communication Constraints: In vector state MIMO LQG, the infinite-horizon cost is $J = \lim_{N \to \infty} J_N \leq p$, tightly relating communication practices (e.g., signals carrying messages) to quadratic optimal control cost (Jahangiri et al., 20 Dec 2025).
• Integrated Sensing–Communications–Control Inequality: In closed-loop industrial systems, a sufficient condition for mean-square convergence (Lyapunov), $E[V(X_{t+1}) | X_t] \leq \rho V(X_t)$, directly links quantization (sensing), communication QoS (delay, bandwidth), and control gain $K$ (Meng et al., 2024).
• Semantic Scheduling with Data Significance: For multi-loop networked control on shared wireless channels, optimal scheduling minimizes weighted MSE given process significance profiles $s_g(\tau)$ (how estimation error amplifies with AoI), capturing the value of control-oriented communication (Roth et al., 2023).

3. Algorithmic Frameworks and Solution Methods

ICAC has spurred the development of multi-agent RL algorithms, co-design heuristics, and convex programming approaches, tailored to the integrated challenges.

• Multi-Agent Hybrid Proximal Policy Optimization with Action Masking (MAHPPO-AM): For UAV swarms, the system is cast as an MDP with hybrid discrete-continuous action spaces (e.g., assignment, power, 3D velocity). Action masking rigorously prunes infeasible actions in high-dimensional spaces, enforcing all assignment and protocol constraints at inference. The actor network includes dedicated branches per action type, and the surrogate PPO loss is clipped with entropy regularization for exploration (Sun et al., 28 Sep 2025).
• Semidefinite Programming (SDP) for LQG Control–Communication Capacity: Communication rate under LQG cost is obtained by maximizing the mutual/directed information from controller to observer over all stationary linear policies, subject to a quadratic cost constraint. This yields a tractable SDP involving innovation covariances and Riccati solutions (Jahangiri et al., 20 Dec 2025).
• Event-Triggered Consensus with Graph Information Bottleneck (CDE-GIB): In MARL with partial observability, agents propagate only messages with high collective significance as determined by a variational IB. A variable-threshold triggers message transmission, further reducing communication without degrading consensus or return, as established by ablation studies (Wang et al., 14 Feb 2025).
• Hybrid-Reward Multi-Agent DRL: For time-sequence semantic control (e.g., robotic teleoperation), distinct agents optimize transmission triggering (when to send/predict) and control gain adaptation, driven by task and communication-oriented rewards. This enables deep reductions in duty cycle with minimal control loss (Li et al., 6 May 2025).
• Active Inference Frameworks (AIF): Joint Bayesian inference and control via free-energy minimization, unifying perception, control, and resource allocation under a probabilistic generative model. Here, both Kalman filtering and action/resource selection (e.g., bandwidth, number of subcarriers) are obtained by sequentially minimizing variational and expected free energy, yielding substantial cost savings compared to separated designs (Pan et al., 17 Sep 2025).

4. Resource Allocation, QoS, and Design Insights

ICAC design necessitates principled allocation and tuning of sensing, computation, and communication resources under control-system requirements.

Parameter	ICAC Influence	Illustrative Work
Bandwidth split	Uplink vs. downlink for queue stability and delay-constrained reliability; more uplink assigned when quantization level $r$ grows; more downlink to meet low packet-loss for tight convergence (Meng et al., 2024)	(Meng et al., 2024)
Energy constraint	Drives trajectory/assignment co-design; optimal policy prevents unnecessary movement and power, directly tuning both throughput and fairness (Sun et al., 28 Sep 2025)	(Sun et al., 28 Sep 2025)
Semantic transmission	MI-based triggering avoids redundant control updates, yielding up to 85% communication reduction at constant tracking error (Li et al., 6 May 2025), significance-aware scheduling yields up to 30% MSE reduction over round-robin (Roth et al., 2023)	(Roth et al., 2023, Li et al., 6 May 2025)
Event-trigger threshold	Variable, decaying thresholds guard against both redundant and stale information (Wang et al., 14 Feb 2025)	(Wang et al., 14 Feb 2025)

• Quantization Rate: Too small $r$ destabilizes estimation, too large induces queuing/backlog delays; optimal $r$ typically in $[5,7]$ bits for practical AGV plants.
• Delay Tradeoffs: Too low $D_{c,\max}$ (allowed closed-loop delay) increases loss rates, too high makes control stale—intermediate values minimize cost (Meng et al., 2024).
• Multi-objective tuning: Closed-form inequalities express tradeoffs among estimation error, bandwidth, and control gains, making parameter sweeps and DE-based global optimization directly interpretable (Meng et al., 2024).

5. Extensions: Over-the-Air and Semantic ICAC Paradigms

• Over-the-Air Control and Sensing: OTA control for UAV swarms shifts control fusion to the BS, drastically reducing the required spectrum usage (all UAVs on one shared band) while simultaneously supporting radar sensing through careful uplink/downlink post-processing and SNR-constrained resource allocation. Closed-form solutions for fusion/dispatch optimize both control and sensing tasks (Wei et al., 11 Feb 2025).
• ISC³ Architectures: Time-sequence-based semantic control—integrating mutual information estimation, LSTM-based reconstruction, and adaptive neural-control policies—jointly optimizes “when to telegraph” and “how much to react,” eliminating unnecessary communication (Li et al., 6 May 2025).
• Active Inference: AIF-based ICAC supports stochastic optimal control with joint sensing and communication resource allocation—minimizing a free-energy functional produces tradeoffs between sensing effort (bandwidth/subcarrier allocation) and closed-loop control performance (Pan et al., 17 Sep 2025).

6. Theoretical Results and Practical Achievability

• Zero Cost-Penalty for Communication: In LQG systems, it is possible to transmit a nonzero rate message over the control channel without elevating the quadratic cost above the pure-control optimum—thus strict separation between control and communication is suboptimal (Jahangiri et al., 20 Dec 2025).
• Semantic Co-Design Achievability: Significance-aware coordination approaches the MSE lower bound, particularly in unstable or critical-loop systems, far outperforming time-driven or distributed random-access approaches (Roth et al., 2023).
• Empirical Performance: RL-based ICAC for UAV swarms achieves fairness index ≈0.99 while reducing energy consumption by ≈25% vs. non-co-design baselines (Sun et al., 28 Sep 2025); event-triggered MARL with information bottleneck reduces communication volume and latent size without sacrificing episodic return (Wang et al., 14 Feb 2025); ISC³ attains ≈85% duty cycle reduction at zero tracking loss (Li et al., 6 May 2025).

7. Limitations, Open Challenges, and Future Directions

Current approaches often depend on Gaussian and linearization assumptions or require heuristic global optimization for non-convexity. Scaling theoretical guarantees to extremely large agent populations ($N \gg 50$), multi-modal latent inference, and hardware-in-the-loop deployment remain pressing. Extensions to semantic, multi-hop, or learning-adaptive ICAC frameworks—potentially with active inference or distributed factor-graph techniques—are indicated as directions for advanced 6G systems (Pan et al., 17 Sep 2025, Wang et al., 14 Feb 2025, Meng et al., 2024). Hardware validation and online adaptation of significance or event thresholds are highlighted as major challenges for robust real-world integration.