Safety-Guaranteed & Goal-Oriented Semantic Communication

• Safety-Guaranteed and Goal-Oriented Semantic Communication is a framework that focuses on transmitting task-critical, semantic information engineered to meet explicit safety and efficiency requirements.
• It employs diverse representations—from scene graphs to logic clauses—and innovative architectures to balance bandwidth, latency, and reliability in complex systems.
• Empirical studies and formal models reveal tradeoffs between semantic abstraction and task preservation while emphasizing the need for certified, end-to-end safety guarantees.

Safety-guaranteed and goal-oriented semantic communication denotes a class of communication systems in which transmitted content is selected, represented, and protected according to its contribution to a downstream task and, in stronger formulations, according to explicit safety requirements on sensing, coordination, or control. Unlike bit-oriented transmission, the central object is not faithful source reconstruction alone but the preservation of task-relevant semantics—object-level evidence, scene graphs, logical clauses, state estimates, or command-and-control data—under bandwidth, latency, and reliability constraints. The contemporary literature establishes a substantial theory of goal-oriented communication, but it also shows that formal safety guarantees remain uneven: many systems provide empirical robustness or constrained average performance, whereas a smaller subset derives task-specific safety conditions or enforces coordination semantics in hardware (Strinati et al., 2020, Zhou et al., 2022, Li et al., 2023).

1. Conceptual foundations

The modern formulation begins with the Shannon–Weaver distinction among the technical, semantic, and effectiveness problems. In this view, classical communication optimizes symbol transport, whereas semantic communication concerns recovery of meaning and goal-oriented communication concerns whether received information enables the intended task or action to be successfully completed with acceptable resource use. A recurring formulation in the 6G literature is that communication should focus not only on “how” to transmit, but also on “what” to transmit, with relevance defined by the source’s intent, the destination’s objective, or the task itself (Strinati et al., 2020, Zhou et al., 2022).

A central conceptual divide runs through the field. One line treats semantics as meaning-bearing structure; another treats semantics as significance or usefulness. The surveys on engineering semantic communication and goal-oriented communication make this distinction explicit: significance-based communication prioritizes freshness, value, relevance, and effectiveness, and context-based communication formulates semantic design as a constrained optimization problem over goals and context (Wheeler et al., 2022). This distinction matters for safety-critical systems because the decisive variable is often not linguistic or perceptual meaning in isolation, but which fragments of information are necessary to avoid hazardous actions.

The Goal-oriented Tensor (GoT) framework makes this shift precise by organizing the communication problem around the triple $\langle X(t), \hat{X}(t), \Phi(t)\rangle$, where $X(t)$ is the true semantic status, $\hat{X}(t)$ is the receiver-side estimate, and $\Phi(t)$ is a context or weighting factor. Its tensor entry,

$$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$

encodes state-context cost, decision gain, and action cost. The framework is explicitly motivated by asymmetric consequences of errors, including the autonomous-driving example in which a false negative can be much more severe than a false positive (Li et al., 2023). That asymmetry is one of the clearest links between goal-oriented semantics and safety-aware design.

2. Formal models and task-aware metrics

Several formal traditions coexist. A foundational information-theoretic line models goal relevance through sufficient statistics and the information bottleneck. In the 6G semantic-communication vision, a task-relevant representation $\mathbf Z=\mathbf f(\mathbf X)$ is sought such that $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$, and the goal-oriented design objective is

$$\min_{\mathbf f(\mathbf X)} I(\mathbf X;\mathbf Z)-\beta I(\mathbf Z;\mathbf \Theta),$$

thereby trading retained source information against information useful for the target variable or task (Strinati et al., 2020).

A concrete task-space formulation appears in goal-oriented wireless image transmission for object detection. There, semantics are induced by a task function $h(\cdot)$, and the semantic discrepancy is not pixel distortion but

$E(X,Y)=\frac{\|h(X)-h(Y)\|}{\|h(X)\|}.$

Communication efficiency is measured by

$X(t)$0

and the combined rate–task tradeoff by

$X(t)$1

The resulting constrained optimization,

$X(t)$2

subject to expected gain and expected error thresholds, exemplifies task-aware communication while also illustrating a recurring limitation: expected constraints are posed, but formal feasibility and worst-case guarantees are not established (Safaeipour et al., 2024).

A more explicitly symbolic line appears in logical decision-making. There, world states are represented by first-order-logic constituents, and task rules induce goal-oriented states

$X(t)$3

Semantic mutual information is defined over logical hypotheses rather than raw symbols, and the goal-oriented selection principle becomes

$X(t)$4

with $X(t)$5 the goal-state space and $X(t)$6 the transmitted evidence. This formulation is notable because it makes clause selection transparent and logically verifiable at the decision level (Saz et al., 21 Apr 2026).

Other metrics generalize task value beyond distortion. In pull-based status updating, the Grade of Effectiveness is

$X(t)$7

combining freshness and usefulness, and is embedded in a cumulative prospect theory objective to model risk awareness and loss aversion (Agheli et al., 9 Mar 2025). In closed-loop sensing–communication–control systems, a hierarchical semantic formulation distinguishes three semantic levels: L1 for observation reconstruction, L2 for state estimation, and L3 for control cost, with separate losses for each level and explicit rate adaptation under a long-term average budget (Pan et al., 22 Dec 2025). Taken together, these models show that “semantic metric” in this area may mean semantic similarity, logical information, control cost, task accuracy, or risk-weighted usefulness, depending on the application.

3. Architectures and semantic representations

A salient feature of the literature is that semantic communication is not tied to a single representation class. In object-detection-oriented image transmission, the semantic bottleneck may be text captions or object-only visual content; the architecture is autoencoder-inspired, with transmitter-side extraction $X(t)$8 and receiver-side reconstruction or interpretation $X(t)$9, but the semantic domain itself remains interpretable rather than a generic latent vector (Safaeipour et al., 2024). This differs sharply from fully opaque end-to-end JSCC.

For wireless visual question answering, the transmitted semantics are question-conditioned visual abstractions. The proposed framework contains six modules—a knowledge base, keywords extractor, question parser, semantic extractor, semantic ranker, and answer reasoner—and instantiates two semantic representations: BBox-based semantics for object-centric questions and scene-graph triplets for relation-heavy questions. The defining property is that semantic selection is conditioned on the current question rather than on generic dataset statistics (Liu et al., 2024).

Road-scene communication for connected autonomous vehicles adopts a scene-graph representation as the primary communication object. The semantic encoder maps an image $\hat{X}(t)$0 into an adjacency tensor $\hat{X}(t)$1 and node-feature matrix $\hat{X}(t)$2,

$\hat{X}(t)$3

with scene relations stacked into $\hat{X}(t)$4. Compression is achieved by removing empty relation slices and reconstructing a task-relevant graph at the receiver rather than raw imagery, after which a GNN + LSTM + MLP pipeline performs downstream risk prediction (Ribouh et al., 9 Mar 2026).

Logical decision-making pushes interpretability further by transmitting subsets of first-order-logic evidence extracted from local perception and reasoning over goal-oriented states induced by task rules (Saz et al., 21 Apr 2026). At the network-systems level, AI-native 6G proposals add semantic infrastructure around these representations. The 6G-GOALS approach introduces a semantic engine at the service-management layer, a semantic RIC for semantic xApps, and a semantic plane spanning CU, DU, and RU, with a shared knowledge base supporting semantic model training, validation, decoding, and interpretation (Strinati et al., 2024). A distinct architectural branch moves from semantic coding to semantic coordination: the TB-CSPN-based FPGA architecture places selected coordination semantics—temporal synchronization, semantic gating, authorization barriers, threshold-based firing, and bounded windows—directly in hardware, separating adaptive software reasoning from deterministic interaction management (Borghoff et al., 2 Jul 2026).

4. Safety assurance and the meaning of “guarantee”

Within this literature, “safety-guaranteed” is not used uniformly. This suggests a three-way distinction. First, some works derive task-specific safety conditions. Second, some enforce coordination constraints architecturally, for example through hardware-level gating and bounded synchronization. Third, many works remain safety-aware or safety-motivated but offer only empirical performance under fixed scenarios. The distinction is critical because expected thresholds, average semantic fidelity, or strong empirical accuracy do not by themselves constitute formal safety certification (Ribouh et al., 9 Mar 2026).

The clearest task-level safety derivation appears in ISAC-enabled UAV obstacle avoidance. There, a Kalman filter supplies position estimates and covariance, and MD-DWA uses Mahalanobis distance to define safe candidate trajectories under uncertainty. The paper derives the mathematical expression of the minimum Mahalanobis distance required to guarantee collision avoidance, using a Gaussian relative-position model, a chi-square confidence region, and a Minkowski-sum construction with the safety disk. Under the stated model, the resulting condition is a conservative sufficient condition for collision avoidance at a $\hat{X}(t)$5 confidence setting (Liu et al., 2 Mar 2026).

A different assurance style is architectural rather than probabilistic. Hardware-enforced semantic coordination places selected TB-CSPN coordination mechanisms directly in FPGA logic to enforce temporal synchronization, semantic gating, authorization constraints, threshold-based collective decisions, bounded time windows, and bounded coordination behavior. Here the strongest guarantee is that selected coordination rules become deterministic and non-bypassable at the hardware layer, while semantic reasoning remains software-driven. The work is explicit, however, that full proofs, invariants, timing bounds, and prototype measurements are not provided in that paper (Borghoff et al., 2 Jul 2026).

Robotics-oriented safety frameworks broaden the notion of what must be guaranteed. The wirelessly connected robotics framework identifies practical safety requirements including safety distance, safety tracking error, safety grasping force, safety speed, and quasi-dynamic stability, and argues that semantic sensing, semantic communication, and semantic control should be organized around these constraints rather than around effectiveness alone (Wu et al., 13 Mar 2026). This is complemented by the security and multi-agent literature. The SemComNet survey organizes risks by control layer, semantic transmission layer, and cognitive sensing layer, emphasizing that unsafe behavior can arise from knowledge-base poisoning, KB mismatch, semantic adversarial attack, semantic relay manipulation, false data injection, or trust failure (Guo et al., 2024). The secure-SemCom survey makes the same point lifecycle-wide: training, model transfer, and semantic information transmission must all be protected because any compromise can alter meaning, not merely packet correctness (Meng et al., 1 Jan 2025).

5. Representative systems and empirical evidence

Wireless image transmission for object detection demonstrates the basic goal-oriented tradeoff. In the text-based, limited-data-rate system, the paper sets $\hat{X}(t)$6 and $\hat{X}(t)$7, and reports average gain $\hat{X}(t)$8, average error $\hat{X}(t)$9, and average weighted error $\Phi(t)$0. In the object-extraction, low-error-tolerance system, it sets $\Phi(t)$1 and $\Phi(t)$2, with average gain $\Phi(t)$3, average error $\Phi(t)$4, and average weighted error $\Phi(t)$5. The result is a direct illustration of the tradeoff between semantic abstraction and task preservation: text semantics maximize communication savings, whereas object-only transmission preserves more task-relevant detail (Safaeipour et al., 2024).

Wireless VQA shows the same principle in a richer multimodal setting. Question-conditioned semantic ranking with BBox or scene-graph representations improves answering accuracy by up to $\Phi(t)$6 under AWGN channels and $\Phi(t)$7 under Rayleigh channels while reducing total latency by up to $\Phi(t)$8 compared to traditional bit-oriented transmission. The gains are most pronounced at lower SNRs, and GO-SG generally outperforms GO-BBox on relation-heavy question types because it preserves relational structure (Liu et al., 2024).

Connected-autonomous-vehicle communication supplies a stronger safety-critical use case. For 5060 images, GBSED reduces raw RGB data from $\Phi(t)$9 GB to $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$0 MB, yielding compression ratio $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$1 and data reduction $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$2. Under the 3GPP CDL channel, semantic fidelity exceeds $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$3 at $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$4 dB and exceeds $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$5 for SNR $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$6 dB. The downstream risk classifier reports Accuracy $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$7, Precision $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$8, Recall $$[\mathcal{C}_1(X(t),\Phi(t))-\mathcal{C}_2(\pi(\hat{X}(t)))]^+ + \mathcal{C}_3(\pi(\hat{X}(t))),$$9, F1-score $\mathbf Z=\mathbf f(\mathbf X)$0, AUC $\mathbf Z=\mathbf f(\mathbf X)$1, and MCC $\mathbf Z=\mathbf f(\mathbf X)$2. The paper interprets this as empirical support for scene-graph communication in safety-critical vehicular tasks, while also stating that no formal collision-probability or missed-hazard guarantees are derived (Ribouh et al., 9 Mar 2026).

Closed-loop control studies reinforce that the choice of semantic level matters. In rate-limited SCC, the L3 GRU-AE achieves the lowest LQR cost among compressed schemes, only $\mathbf Z=\mathbf f(\mathbf X)$3 higher than the rate-unconstrained LQR benchmark. With an average communication budget of 4 bits, L3 adaptive allocation has $\mathbf Z=\mathbf f(\mathbf X)$4 performance loss relative to uncompressed communication but reduces LQR cost by $\mathbf Z=\mathbf f(\mathbf X)$5 and $\mathbf Z=\mathbf f(\mathbf X)$6 compared with L1 and L2 schemes, respectively, showing that control-oriented semantics is distinct from reconstruction-oriented semantics (Pan et al., 22 Dec 2025). In pull-based status updating, effect-aware scheduling under strict cost constraints yields average CPT-based total GoE that is $\mathbf Z=\mathbf f(\mathbf X)$7 higher for the model-based method and $\mathbf Z=\mathbf f(\mathbf X)$8 higher for the model-free method than LWGF at $\mathbf Z=\mathbf f(\mathbf X)$9, indicating that freshness and usefulness can be jointly optimized under explicit query-cost constraints (Agheli et al., 9 Mar 2025).

The strongest end-to-end safety-oriented robotic evidence comes from UAV obstacle avoidance and target tracking. In ISAC-enabled obstacle avoidance, the GOSC framework achieves the same $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$0 task success rate as conventional always-on transmission while reducing the number of transmitted sensing and command-and-control signals by $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$1 and the number of transmission time slots by $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$2 (Liu et al., 2 Mar 2026). In wirelessly connected UAV target tracking, semantic-based command execution improves safety rate and tracking success rate by more than $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$3 times and $$I(\mathbf X;\mathbf \Theta)=I(\mathbf Z;\mathbf \Theta)$$4 times, respectively, relative to the baseline, especially under stricter safety constraints (Wu et al., 13 Mar 2026).

6. Limitations, misconceptions, and future directions

A persistent misconception is that constrained task-aware optimization is equivalent to a safety guarantee. The literature repeatedly rejects that reading. Expected gain/error constraints in goal-oriented image transmission are design targets rather than proofs; empirical risk-assessment performance in autonomous driving is not a certified missed-hazard bound; and many robotics frameworks use “safety-guaranteed” as a design objective while remaining light on control-theoretic proof machinery (Safaeipour et al., 2024, Ribouh et al., 9 Mar 2026, Wu et al., 13 Mar 2026). Another misconception is that semantic communication is necessarily an opaque neural latent-code paradigm. Current systems also use captions, object crops, BBoxs, scene graphs, first-order-logic clauses, and hardware-level semantic tokens (Liu et al., 2024, Saz et al., 21 Apr 2026).

Open challenges are correspondingly structural. The GoT framework identifies communication networks with heterogeneous goals, goal-oriented physical-layer techniques, and perception–communication–computation–control co-design as unresolved problems (Li et al., 2023). AI-native 6G architectures add semantic mismatch, timing-aware distributed reasoning, causal semantic representation, and coexistence with legacy systems (Strinati et al., 2024). Multi-agent surveys add KB synchronization, semantic interoperability, explainable semantic models, endogenous security, and trust management as prerequisites for dependable SemComNet deployment (Guo et al., 2024).

A plausible research trajectory, explicitly suggested in several papers, is to move from average-case effectiveness toward safety-first constraints. Examples already proposed in the literature as future extensions include constrained risk minimization over safety-critical classes, chance constraints on miss-detection probability, worst-case or distributionally robust optimization, confidence-aware transmission with fallback to richer data, certified robustness, channel-aware unequal error protection for critical semantics, control barrier functions, robust or tube MPC driven by decoder uncertainty, and lexicographic safety-first scheduling of semantic updates (Safaeipour et al., 2024, Pan et al., 22 Dec 2025, Agheli et al., 9 Mar 2025). Logical and hardware-based approaches suggest complementary routes: the former offers clause-level verifiability, while the latter offers deterministic enforcement of interaction rules (Saz et al., 21 Apr 2026, Borghoff et al., 2 Jul 2026). The field’s central unresolved problem is therefore not whether communication can be made goal-oriented—that is already well established—but how to attach formal, end-to-end, safety-relevant guarantees to semantic representations, schedulers, decoders, and controllers without forfeiting the efficiency gains that motivated semantic communication in the first place.