Communication-Control Enhanced Sensing

• Communication-and-Control-Enhanced Sensing is a unified design paradigm that integrates sensing, communication, and control to optimize information flow in dynamic, resource-constrained environments.
• It leverages bidirectional feedback and adaptive decision-making—using techniques like reinforcement learning and Kalman filtering—to enhance control accuracy and resource efficiency.
• Empirical studies demonstrate significant gains such as up to an 85% reduction in communication duty cycles and improved latency, validating its efficacy in complex cyber-physical scenarios.

Communication-and-Control-Enhanced Sensing (CCES) refers to a systemic paradigm in which sensing, communication, and control are jointly co-designed and tightly coupled within feedback loops to achieve superior closed-loop performance in autonomous, robotic, and cyber-physical systems. Unlike traditional architectures, which treat sensing, communication, and control as isolated functionalities, CCES leverages bidirectional information flow, semantic data selection, and feedback from control objectives to optimize what, when, and how information is sensed and transmitted. This unified approach enables enhanced perception accuracy, resource efficiency, and robust performance in complex or resource-constrained environments.

1. Principles and Motivations

CCES emerges from the recognition that harsh environments, limited resources, and stringent control requirements demand that sensing, communication, and control subsystems do not operate in isolation. Classical designs involve periodic or event-based sensing, fixed communication schedules, and open-loop estimation; these approaches are suboptimal when bandwidth is limited, environmental dynamics are variable, and actuation depends sensitively on information quality. By establishing robust feedback loops whereby control tasks actively inform sensing strategies and communication resource allocations, CCES ensures that only the most relevant, task-critical information is acquired, encoded, and delivered—thereby reducing overall latency, data load, and energetic cost while maintaining or improving end-to-end system reliability and control accuracy (Zhao et al., 8 Oct 2025, Cai et al., 14 Feb 2025, Wei et al., 21 Jan 2026).

The main drivers for CCES include:

• Dynamic adaptation to environmental and control state, improving robustness under fading, interference, or varying geometry (Zhao et al., 8 Oct 2025, Wang et al., 11 Aug 2025).
• Resource optimization (transmission energy, bandwidth, sensing time) through selective sampling and communication, guided by control performance gradients (Cai et al., 14 Feb 2025, Wei et al., 21 Jan 2026).
• Enhanced stability, tracking, and estimation in networked cyber-physical systems by closing the sensing–communication–control loop (Wei et al., 21 Jan 2026, Meng et al., 2024).

2. Mathematical Models and System Architectures

CCES systems are typically structured as multi-block feedback architectures integrating:

• Sensing Layer: At time $t$, the system observes raw or processed measurements (e.g. images $w_t$, state vectors $x_t$).
• Communication Layer: Observations are encoded into channel symbols $s_t$ using adaptive semantic encoders informed by real-time control relevance (Zhao et al., 8 Oct 2025, Li et al., 6 May 2025).
• Control Layer: Received (possibly distorted) measurements are used to update belief states $b_t$ and generate control actions $u_t$, which in turn inform future sensing/communication policies.

A canonical instance is the lunar landing scenario (Zhao et al., 8 Oct 2025):

$$\begin{align*} \text{(Sensing)}\quad & w_t \rightarrow y_t = \eta(w_t)\ \text{(Semantic Encoding)}\quad & s_t = E_e(w_t; \alpha, \delta_t) \ \text{(Channel)}\quad & \hat{s}_t = h(t) s_t + n \ \text{(Decoding)}\quad & \hat{w}_t = E_d(\hat{s}_t; \alpha, \delta_t)\ \text{(Semantic Extraction)}\quad & z_t = \eta(\hat{w}_t)\ \text{(Belief Update)}\quad & b_t = \tau(b_{t-1}, u_{t-1}, z_t)\ \text{(Control)}\quad & u_t = \pi(b_t) \end{align*}$$

Here, $\delta_t$ denotes the fraction of high-priority semantic features transmitted, dynamically adjusted according to the control sensitivity.

Across recent literature, variants of this architecture integrate Kalman or Bayesian filters, semantic mutual information extractors, adaptive quantizers, and resource allocation blocks, sometimes augmented by reinforcement learning for dynamic adaptation (Li et al., 6 May 2025, Cai et al., 14 Feb 2025, Wei et al., 21 Jan 2026).

3. Adaptive Sensing and Control Relevance

A central innovation in CCES is the prioritization of sensory information according to its real-time impact on control performance. This is often formalized using control-centric metrics such as the gradient of reward with respect to control input, $\partial r / \partial u_t$, or information-theoretic utility. For example, in (Zhao et al., 8 Oct 2025), image transmission is dynamically focused on those patches that are most critical for accurate control in high-sensitivity regimes, as scored by a dynamic vision transformer (DynamicViT) backend.

The adaptation loop includes:

• Feature importance scoring: $\zeta_{tk}$ for each measurement, selecting top-$\delta_t N$ patches for full-resolution encoding.
• Control-aware decision: $$\delta_t = 1/(1+\exp[\kappa \cdot \partial r(b_t,u_t)/\partial u_t])$$, increasing resource allocation when the system is near critical maneuvers (Zhao et al., 8 Oct 2025).
• Sensing acquisition is terminated or intensified based on online variance reduction or real-time model feedback (e.g., gradient importance sampling in AI-in-the-loop frameworks (Cai et al., 14 Feb 2025)).
• Semantic feature extractors and reconstructors (e.g., MI neural networks, LSTMs) to minimize redundant transmissions while filling in skipped intervals at the control node (Li et al., 6 May 2025).

This control-aware sensing strategy leads to substantial reductions in communication and sensing load without degradation of closed-loop control performance, and is robust to deep communication fades or environmental disruptions.

4. Joint Optimization and Solution Approaches

The end-to-end (E2E) co-design problem is inherently multiobjective: minimize closed-loop LQR or RL-based control cost, sensing estimation error, communication delay, and energy, subject to QoS and stability constraints. The optimization over control laws, sensing allocation, and communication scheduling is typically non-convex and high-dimensional. Representative strategies include:

• Alternating Optimization: Block-coordinate descent between control, sensing, and communication variables (e.g., ADMM variants as in (Wei et al., 21 Jan 2026)). Subproblems are convexified via surrogate functions (e.g., SCA, SDP relaxations) (Wang et al., 11 Aug 2025, Jin et al., 26 Jun 2025).
• Semantic Communication Co-design: End-to-end loss functions penalize both reconstruction error and task-inferred control cost (Zhao et al., 8 Oct 2025, Li et al., 6 May 2025).
• Reinforcement Learning: Adaptive control-gain and update decision policies are trained via deep RL, leveraging multi-agent architectures in time-sequence-based semantic ISCC (Li et al., 6 May 2025).
• Self-triggered Control: Sparse activation of sensing and communication via optimal selection of measurement times and feedback structure to ensure a guaranteed cost bound (Bahavarnia et al., 2018).
• Active Inference: Joint state estimation, resource-aware sensing, and control using variational and expected free energy minimization, leading to Kalman-like estimation coupled with dynamic subcarrier allocation (Pan et al., 17 Sep 2025).

Constraints are enforced for system stability (Lyapunov or Riccati-based), maximum allowable delay or loss, and bandwidth/energy budgets (Meng et al., 2024, Jin et al., 26 Jun 2025, Meng et al., 2023).

5. Performance Gains and Empirical Evidence

Quantitative studies consistently demonstrate that CCES schemes outperform both traditional separated pipelines and static ISCC baselines in a range of benchmarks:

• Robustness to Channel Fading: Semantic, control-aware coding outperforms JPEG and non-adaptive pipelines in deep fades, maintaining high PSNR and crater-matching in lunar landing images at low SNR (Zhao et al., 8 Oct 2025).
• Resource Efficiency: Up to 85% reduction in communication duty cycle without loss of control tracking (teleoperation, time-sequence semantic communication (Li et al., 6 May 2025)); up to 77% energy savings and 52% lower sensing cost in federated learning tasks using AI-in-the-loop JSAC (Cai et al., 14 Feb 2025).
• Tracking and Latency: Integrated SCEC in UAV swarms reduces tracking RMSE by 35% and mean comm latency by 40% versus separate control/sensing allocations (Wei et al., 21 Jan 2026).
• Industrial Control Systems: Joint tuning of quantization, bandwidth, and control gain prevents queue overflow and guarantees fast, stable convergence, even under tight bandwidth and delay budgets (Meng et al., 2024, Meng et al., 2023).
• Agility in Mobility Scenarios: Adaptive co-design in MANETs and leader-follower robotics (including optical sensing/comm) enables robust, rapid reaction to abrupt maneuvers and environmental variability (Wang et al., 2023).
• Enhanced Generalization: AI-in-loop frameworks achieve lower validation loss in distributed model training under energy and latency constraints (Cai et al., 14 Feb 2025).

A summary of selected empirical results is shown below:

Study	Main Gain (vs baseline)	Metric/Scenario
(Zhao et al., 8 Oct 2025)	+50% crater-matching	Lunar landing under Rician fade
(Cai et al., 14 Feb 2025)	–77% com energy, –58% loss	Federated edge learning
(Wei et al., 21 Jan 2026)	–35% RMSE, –40% latency	UAV swarm formation/hybrid SCEC
(Li et al., 6 May 2025)	–85% duty cycle, <1mm RMS	Real robot teleoperation/semantic SI
(Meng et al., 2024)	“U-shaped” optimal trade-off	AGV control, resource-constrained

6. Implementation Domains and Extensions

CCES methodologies are being deployed in trajectories across:

• Lunar/Planetary Landing and Remote Robotics: Real-time semantic image transmission and control amid extreme channel impairments (Zhao et al., 8 Oct 2025).
• Swarms and Multi-agent Systems: ADMM and distributed Kalman-fusion based co-optimization in UAV swarms, formation flight, and cooperative ground robots (Wei et al., 21 Jan 2026, Jin et al., 26 Jun 2025, Wang et al., 2023).
• Industrial Automation and IIoT: 5G-URLLC-enabled closed-loop control of AGVs and robotic arms, with joint stability-bounded design of communication, quantization, and sensing schedules (Meng et al., 2023, Meng et al., 2024).
• Edge AI and Learning Systems: Model-driven joint sampling, gradient aggregation, and SGLD-noise adaptation for resource-efficient federated and edge learning (Cai et al., 14 Feb 2025).
• Non-Terrestrial and Optical ISAC: Energy-efficient satellite–UAV–IoT, and non-RF optical integrated communication/sensing (Liang et al., 2024, Wang et al., 2023).
• Active Inference and Bayesian Approaches: Free-energy based closed-loop resource allocation and state estimation in communication-constrained UAVs (Pan et al., 17 Sep 2025).

Further extensions include integration with 6G service fabrics (DS3C), intent-based service APIs, multi-agent RL for hybrid control and communication policy adaptation, and robustification against adversarial or non-Gaussian noise (Vukobratović et al., 2023).

7. Limitations, Challenges, and Future Directions

Despite significant advances, key open challenges remain:

• Scalability and Real-time Computation: Convexification and decomposition methods are critical for real-time deployment but may introduce suboptimalities; high-dimensional, low-latency CCES in large swarms or industrial environments remains a challenge (Wang et al., 11 Aug 2025, Wei et al., 21 Jan 2026).
• Generalization beyond Linear-Gaussian Models: Many analyses assume linear plant dynamics or Markovian structures; extensions to highly nonlinear, non-Markovian, or partially observed systems are ongoing (Li et al., 6 May 2025, Pan et al., 17 Sep 2025).
• Heterogeneous and Nonstationary Networks: Handling diverse channel models, mobility, and interference in dynamic networked environments requires further refinement of adaptive and learning-based CCES policies (Zhao et al., 8 Oct 2025, Liang et al., 2024).
• Integration with Standardization and Orchestration: Embedding CCES within 6G distributed fabrics, leveraging intent-based orchestration APIs, and cross-domain resource negotiation raise new architectural issues (Vukobratović et al., 2023).
• Security and Privacy: Control- and communication-aware sensing introduces new vectors for adversarial manipulation if not robustly designed.

A plausible implication is that future cyber-physical and robotic platforms will mandate tightly integrated CCES stacks, bridging AI, wireless, and control-theoretic disciplines to achieve resilient, resource-efficient, and high-precision autonomous operation in both terrestrial and extra-terrestrial contexts.