Joint Communication and Control

• Joint Communication and Control (JCC) is the integrated optimization of wireless communication and control variables where impairments directly affect closed-loop system performance.
• It balances trade-offs between transmission delay, stability, and control error through analytical modeling, optimization algorithms, and learning-based strategies.
• JCC spans applications in vehicular platoons, industrial cyber-physical systems, AGVs, and UAVs, showcasing improvements in reliability and task performance.

Joint Communication and Control (JCC) denotes a class of co-design problems in which communication variables—such as access, association, scheduling, power allocation, coding, or beamforming—and control variables—such as feedback gains, actuation policies, trajectories, or stability margins—are optimized jointly because wireless impairments enter directly into the closed loop. In wireless networked control systems, vehicular platoons, edge-assisted industrial cyber-physical systems, and embodied-intelligence architectures, the objective is not merely reliable transmission but the joint optimization of communication performance and control outcomes such as stability, latency, task error, and Value of Information (VoI) (Zeng et al., 2018, Liu et al., 19 Jan 2025, Gan et al., 9 Apr 2026, Lei et al., 11 May 2025). The terminology is not fully uniform: in mobile-edge and antenna-system literature, the same acronym can also denote joint communication and computation, often with closely related resource-allocation and task-execution objectives (Chen et al., 2018, Gan et al., 20 Nov 2025).

1. Coupling mechanisms and system view

The defining premise of JCC is that communication is part of the plant-controller loop rather than an external service layer. In autonomous vehicular platoons, the follower control law explicitly depends on delayed V2V measurements, so wireless delay directly affects plant stability and string stability (Zeng et al., 2018). In edge-assisted wireless control, each subsystem comprises a plant, sensor, and actuator; sensing data are offloaded to a base station, processed by MEC, and returned as command signals, so uplink transmission, edge computation, and downlink broadcasting jointly determine the closed-loop sample period (Liu et al., 19 Jan 2025). In the unified analytical JDCC model, the control loop is explicitly uplink-downlink closed loop, and communication and control coexist through shared resources (Gan et al., 9 Apr 2026).

This system view departs from architectures that optimize networking and control independently. Several works state that prior communication analyses often ignore control stability, while control-centric designs assume ideal or delay-free communication (Zeng et al., 2018). JCC instead treats delay, outage, interference, finite blocklength effects, power limits, computation slots, and mobility as endogenous variables of the control problem.

The same coupling reappears at larger architectural scales. In B2X networks, the “brain” performs reasoning and planning, the “body” senses and acts, and “X” denotes the surrounding ecosystem; uplink state acquisition and downlink command delivery are both redesigned for the embodied loop rather than for generic traffic (Liu et al., 1 Jul 2026). This suggests a broader interpretation of JCC as loop-level co-design across sensing, communication, computing, and control, with communication policies evaluated by their effect on embodied action.

2. Performance metrics and trade-off characterizations

JCC research is organized around coupled performance metrics. In the unified JDCC framework, the two fundamental metrics are communication transmission delay and steady-state control variance (Gan et al., 9 Apr 2026). For a communication payload of $Q_U$ bits and downlink SINR $\Gamma_U$, the transmission delay is

$$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$

For the discrete-time LTI process $x_{n+1} = a x_n + b u_n + w_n$, the steady-state control variance is

$$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$

with stability conditions

$$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$

These expressions yield a Pareto boundary in the $(\tau_U, V_\infty)$ plane and make explicit that communication delay and control error are jointly constrained rather than separately minimized (Gan et al., 9 Apr 2026).

In platoon control, the key metric is whether the wireless system satisfies control-derived delay limits. The plant-stability delay threshold is

$$\tau_1 = \frac{\lambda_{\min}(\mathbf{M}_3)}{\lambda_{\max}(\mathbf{M}_4)},$$

the string-stability threshold is

$\tau_2 = \frac{C^2 - 2A - B^2}{2AC},$

and the maximum tolerable wireless delay is $\min(\tau_1,\tau_2)$ (Zeng et al., 2018). Reliability is then defined as the probability that end-to-end V2V delay does not exceed this maximum delay.

In AGV networked control, communication metrics include delay, packet loss, and Packet Reception Ratio (PRR), whereas control accuracy is quantified by the path-tracking error

$\Gamma_U$0

between planned and actual trajectories (Bragato et al., 8 Sep 2025). In VoI-based formulations, the bridge variable is not a physical-layer KPI but a control-performance differential. In the VDTN framework, the Expected Cumulative VoI is

$\Gamma_U$1

and the Immediate VoI $\Gamma_U$2 measures the control-performance loss induced by acting on impaired observations at control interval $\Gamma_U$3 (Lei et al., 12 May 2025).

Metric	Representative expression	Role
Communication delay	$\Gamma_U$4	Latency in JDCC
Control variance	$\Gamma_U$5	Steady-state control quality
Delay-feasibility reliability	$\Gamma_U$6	Wireless support for stability
Tracking error	$\Gamma_U$7	AGV path accuracy
Expected Cumulative VoI	$\Gamma_U$8	Control-aware valuation of data

A recurring conclusion is that the relevant operating point is a trade-off frontier rather than a single optimum. The Pareto boundary in JDCC, the reliability-versus-delay thresholds in platoons, and the PRR-versus-tracking error relation in AGVs all formalize the same principle: communication quality must be judged by its effect on the closed loop (Gan et al., 9 Apr 2026, Zeng et al., 2018, Bragato et al., 8 Sep 2025).

3. Optimization formulations and algorithmic patterns

The optimization structure of JCC problems is typically non-convex because of mixed discrete-continuous decisions, finite-blocklength reliability constraints, interference terms, or nonlinear dynamics. In multi-BS edge control, the main problem minimizes the closed-loop control latency $\Gamma_U$9 subject to uplink and downlink outage constraints, Lyapunov-based control-stability constraints, computation-resource limits, power constraints, unique association constraints, and the TDMA frame relation

$$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$0

The proposed solution relaxes binary association variables $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$1, alternates between association/time allocation and power/time allocation, applies successive convex approximation (SCA) to non-convex constraints, and finally rounds the relaxed association matrix to a binary assignment (Liu et al., 19 Jan 2025).

Vehicular platoon JCC uses a different decomposition. There, the communication model supplies a probabilistic delay characterization, while the control layer supplies stability-induced delay requirements; the controller gains $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$2 and $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$3 are then optimized to maximize wireless reliability, subject to stability conditions $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$4 and $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$5 (Zeng et al., 2018). Because the resulting optimization is non-convex, a dual-update Lagrangian method is used to obtain sub-optimal gain values.

Learning-based JCC has introduced a separate family of formulations. In nonlinear control-non-affine WNCS, a deep Koopman model learns an embedding in which the dynamics become linear,

$$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$6

and control is computed by an LQR law in the embedding space,

$$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$7

followed by a learned decoder $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$8 (Vijithasena et al., 2 Sep 2025). Communication scheduling is then performed via Lyapunov optimization, with a virtual-queue formulation that trades off transmission frequency and state-prediction error.

VoI-based JCC replaces explicit model coupling by a reward-coupling mechanism. In VDTN, control and communication are decoupled into two interrelated MDPs operating at different time scales, and two DRL modules are trained iteratively: the control module estimates VoI, and the communication module optimizes radio resource allocation using a reward that combines a communication metric and VoI (Lei et al., 12 May 2025). The SSDP framework generalizes this idea by introducing exogenous information $$\tau_U = \frac{Q_U}{B_{\mathrm{dn}}\log_2(1 + \Gamma_U)}.$$9 explicitly into the sequential decision process,

$x_{n+1} = a x_n + b u_n + w_n$0

and uses VoI-based rewards to optimize the “When”, “What”, and “How” to communicate problems (Lei et al., 11 May 2025).

A related pattern appears in mobility-assisted wireless design, where the control variable is the motion of the communication platform. In cognitive UAV communication, average achievable rate is maximized by jointly optimizing UAV trajectory or 3D maneuver and transmit power under UAV motion constraints and interference-temperature constraints at primary receivers, using alternating optimization and SCA (Huang et al., 2018, Huang et al., 2019). Although these works are not control-theoretic in the state-feedback sense, they treat platform maneuver as a control variable coupled to communication performance.

4. Representative application domains

JCC has been instantiated across transportation, industrial automation, edge robotics, and UAV systems (Zeng et al., 2018, Liu et al., 19 Jan 2025, Lei et al., 12 May 2025, Diao et al., 11 Mar 2025, Bragato et al., 8 Sep 2025).

Domain	Joint variables	Reported outcome
Vehicular platoons	V2V delay and control gains	Reliability improves by up to 15%
Multi-BS edge WNCS	Association, power, time, computation	Lower latency and lower control cost than heuristics
Vehicular digital twin networks	DRL control and radio resource allocation	V2I throughput $x_{n+1} = a x_n + b u_n + w_n$1 Mbps vs $x_{n+1} = a x_n + b u_n + w_n$2 Mbps
Industrial CPS in CARLA	Task-oriented JSCC and delay-aware control prediction	Driving score $x_{n+1} = a x_n + b u_n + w_n$3 at 1 s E2E delay
Industrial AGV NCS	Network impairment and predictive path tracking	PRR strongly correlated with tracking error

The platoon literature is one of the clearest demonstrations of closed-loop dependence. With optimized control gains, wireless reliability can improve by up to $x_{n+1} = a x_n + b u_n + w_n$4, and the same reliability target can be met with larger platoons or increased spacing relative to non-optimized control settings (Zeng et al., 2018). In the VDTN platoon scenario, the VoI-guided JOCC framework attains higher average throughput per V2I link, $x_{n+1} = a x_n + b u_n + w_n$5 Mbps versus $x_{n+1} = a x_n + b u_n + w_n$6 Mbps for a delay-aware DRL baseline, without sacrificing control stability (Lei et al., 12 May 2025).

Industrial CPS has pushed JCC toward task-oriented perception and predictive control. In edge-enabled autonomous driving, task-oriented JSCC based on Information Bottleneck is co-designed with Delay-Aware Trajectory-Guided Control Prediction (DTCP); under an end-to-end delay of $x_{n+1} = a x_n + b u_n + w_n$7 second ($x_{n+1} = a x_n + b u_n + w_n$8 time slots), the framework achieves a driving score of $x_{n+1} = a x_n + b u_n + w_n$9, which is $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$0 points higher than using Better Portable Graphics while reducing bandwidth usage by $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$1 (Diao et al., 11 Mar 2025). This is not a generic source-coding result: the transmitted representation is optimized for downstream driving actions rather than for image fidelity.

Simulation-based AGV studies give a complementary picture at the systems level. In a ROS 2 and Gazebo framework, lower PRR produces dramatically larger path deviation, and for $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$2 the control error drops sharply to about $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$3 m, whereas increasing delay from $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$4 to $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$5 ms with perfect PRR raises the control error only from about $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$6 to $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$7 m (Bragato et al., 8 Sep 2025). The same study reports that increasing the look-ahead parameter $$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$8 reduces MSE at low PRR, showing that predictive control can partially compensate for unreliable communication.

UAV-assisted target tracking extends the scope from JCC to integrated sensing, communication, and control. There, a stochastic MPC problem jointly optimizes control inputs and beamforming, target state is estimated with an EKF, and closed-form beamforming under fixed control inputs is used to reformulate the problem into a tractable control-oriented form; numerical results show tracking accuracy comparable to a non-causal benchmark while maintaining stable communication (Chen et al., 5 Feb 2026).

5. Resource-allocation, computation, and economic extensions

A substantial branch of the literature uses the same co-design logic but places communication in direct interaction with computation resources. In on-demand mobile-edge cloud computing, JCC denotes joint communication and computation resource allocation. The framework contains a user-side resource-efficient computation offloading mechanism, operator-side admission control and JCC resource allocation, and a truthful pricing scheme for strategic users (Chen et al., 2018). User tasks are modeled as DAGs, delay-aware graph partitioning is solved by dynamic programming via backward induction, and the operator admission-control problem is shown to be strongly NP-hard. A low-complexity approximation ranks users by

$$V_\infty = \frac{\sigma_w^2 S_\alpha \Gamma_\alpha}{S_\alpha \Gamma_\alpha - |a|^2 (S_\alpha + \Gamma_\alpha - 1)},$$9

and admits users greedily under communication and computation capacities. The paper reports a $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$0-approximation in a homogeneous special case, an $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$1 runtime, empirical performance loss of at most about $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$2 relative to exact solution, and a critical-value pricing rule that guarantees truthfulness (Chen et al., 2018).

The extension from JCC to compute-caching-communication control further broadens the design space. In online data-intensive service delivery, services are modeled by DAGs and an Augmented Layered Graph, and the DI-DCNC policy is shown to be throughput-optimal for joint path selection, processing-location decisions, cache selection, and traffic scheduling (Cai et al., 2022). Reported gains include $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$3–$$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$4 higher maximum sustainable arrival rates than benchmark schemes, up to $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$5 savings in compute and communication resources in some configurations, and roughly doubled effective throughput under dynamic database replacement when storage is limited (Cai et al., 2022).

A different communication-computation branch appears in the segmented PASS framework, where both communication bit streams and computation data are transmitted simultaneously via uplink. Under segment selection, segment aggregation, and segment multiplexing protocols, the paper formulates MSE minimization for computation-oriented cases and WSR maximization for communication-oriented cases, using AO-MMSE and AO-WMMSE algorithms. Reported gains reach $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$6 and $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$7 reductions in MSE relative to conventional MIMO and conventional PASS, and $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$8 and $$S_\alpha > |a|^2,\qquad \Gamma_\alpha > \frac{|a|^2(S_\alpha-1)}{S_\alpha - |a|^2}.$$9 improvements in WSR, respectively (Gan et al., 20 Nov 2025).

These results do not collapse JCC into computation offloading, but they show that many practical closed loops at the network edge are triadic rather than dyadic: command quality depends on communication, and communication feasibility depends on shared compute, caching, or over-the-air aggregation resources.

6. Misconceptions, limits, and emerging directions

A common misconception is that communication and control can be evaluated independently and recombined afterward. The JDCC outage analysis explicitly rejects this simplification: the joint outage probability is not reducible to the product of communication-only and control-only outages because the event $(\tau_U, V_\infty)$0 is coupled through shared resources and the uplink-downlink loop (Gan et al., 9 Apr 2026). Similarly, in multi-BS edge control, basic proximity is not optimal for association; more subsystems may be assigned to a BS with higher power or computational capacity even if it is farther away (Liu et al., 19 Jan 2025).

Another misconception is that delay is always the dominant network impairment. The AGV study indicates a more nuanced picture: packet loss and PRR are the dominant factors for trajectory tracking, whereas delays up to $(\tau_U, V_\infty)$1 ms produce comparatively small degradation under perfect PRR (Bragato et al., 8 Sep 2025). This does not imply that delay is unimportant; rather, it shows that the controlling communication bottleneck is application-dependent.

Emerging work also questions the assumption that control requires explicit reliable message transmission. The Joint Detection and Identification framework for nanorobot swarms shifts from reliable per-device communication to a control-oriented broadcast model in which devices locally decide whether a message is relevant. The paper emphasizes favorable asymptotic scaling for identification, but also states that practical implementations are governed by finite blocklength effects, noise, and latency (Labidi et al., 30 Mar 2026). This introduces a control-layer abstraction in which the central problem is subset activation rather than explicit decoding.

The B2X perspective pushes JCC toward embodied-intelligence networking. It distinguishes distributed-brain and centralized-brain architectures, redesigns the uplink around event-triggered low-latency access, high-volume sensing access, and massive short-packet access, redesigns the downlink around data-command multiplexed delivery, periodic command delivery, and mobility-robust command delivery, and uses a communication-control Pareto boundary to characterize loop-level trade-offs (Liu et al., 1 Jul 2026). The main open problems identified there—loop-level performance modeling, distributed brain placement, and function- and energy-aware protocol co-design—indicate that JCC is evolving from a link-level cross-layer optimization topic into a systems framework for wireless embodied autonomy.

Taken together, the literature presents JCC as a family of formulations rather than a single model class. Some works are analytical and derive Pareto frontiers, outage expressions, or delay thresholds; some are algorithmic and use alternating optimization, SCA, Lyapunov optimization, dynamic programming, or truthful pricing; and some are learning-based and use DRL, SSDP, or deep Koopman embeddings. What remains invariant is the central claim: communication performance is meaningful only relative to the control or task objective that it supports.