Semantic & Task-Oriented V2X Communications

• Semantic and task-oriented V2X communications are vehicular communication paradigms that transmit only essential semantic features needed for tasks like collision avoidance and cooperative perception.
• They employ advanced techniques such as context-aware encoding, joint source–channel coding, and receiver-specific content curation to meet strict bandwidth and latency requirements.
• Empirical studies show substantial gains in collision prediction accuracy, significant data volume reductions, and enhanced network scalability for real-world applications.

Semantic and task-oriented V2X communications denote a class of vehicular communication paradigms in which the transmitted payload is selected, encoded, and scheduled according to its meaning and its utility for a receiver’s driving task, rather than according to bit-level fidelity alone. In this formulation, Vehicle-to-Everything links are designed to deliver compact semantic features, object-level descriptors, predictive embeddings, scene graphs, or trajectory-relevant summaries that are sufficient for collision avoidance, cooperative perception, platooning, hazard communication, map updating, or trajectory prediction under stringent bandwidth and latency constraints. Across recent work, the field is framed as a progression from Shannon–Weaver Level A communication toward Level B semantic communication and Level C task-oriented communication, with context-aware relevance estimation, semantic compression, joint source–channel coding, and receiver-specific content curation as recurrent design principles (Lyu et al., 2024, Lusvarghi et al., 10 Aug 2025, Lusvarghi et al., 8 Jun 2026).

1. Conceptual foundations and formal definitions

The central distinction from conventional V2X lies in the communication objective. Classical systems optimize reliability, throughput, BER, PER, and latency for raw or weakly processed data. Semantic communication instead seeks to transmit only the information needed to convey meaning for a receiver’s inference or control task, while task-oriented communication further restricts transmission to the subset required for successful actuation or decision making. One common formalization introduces a semantic encoder and decoder, $s = \Phi(x; K_{tx})$ and $\hat{s} = \Psi(r; K_{rx})$, where $x$ is a sensor measurement, $K_{tx}$ and $K_{rx}$ are background knowledge, and the receiver either reconstructs semantics or directly performs an action $a = g(\hat{s})$ (Lyu et al., 2024). A related abstraction models a semantic source as a pair $(\mathcal X,\mathcal Z)$ with an encoding $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$, where $\mathcal Z$ may represent, for example, bounding boxes, classifications, and motion vectors (Lusvarghi et al., 10 Aug 2025).

Within this framework, vehicular context is treated as a first-class variable. A connected and autonomous vehicle is described through endogenous state, such as own position, speed, and planned path, and exogenous information derived from sensed or received objects. In a context-aware formulation, all context accessible to vehicle $V_n$ is bundled into a context vector $\hat{s} = \Psi(r; K_{rx})$0, and the relevance of object $\hat{s} = \Psi(r; K_{rx})$1 to that receiver is expressed as $\hat{s} = \Psi(r; K_{rx})$2, often implemented by a discrete contextual relevance function with zero weight for irrelevant or redundant objects (Lusvarghi et al., 10 Aug 2025). Semantic and task-oriented V2X therefore operates not only on source compression, but on receiver-conditioned selection.

A closely related line of work uses the term pragmatic communications, or PragComm, to emphasize “effective communications” rather than full scene reconstruction. In collaborative autonomous driving, this corresponds to transmitting only those semantic features that affect imminent planning and control decisions. The “less is more” principle used in Select2Drive states that a broadened perceptual horizon can confuse the decision module rather than improve it, so communication should be restricted to the Area of Importance around the vehicle’s planned trajectory (Huang et al., 21 Jan 2025). This suggests that semantic and task-oriented V2X is not merely a compression layer over conventional V2X, but a redesign of the message-generation criterion itself.

2. Semantic representations and end-to-end architectures

Recent systems implement semantic V2X with several distinct representation families. One family uses predictive latent embeddings. In "Spatiotemporal Semantic V2X Framework for Cooperative Collision Prediction" (Onsu et al., 23 Jan 2026), an RSU-mounted camera uses V-JEPA to generate spatiotemporal semantic embeddings of future frames. The model is pretrained by masked predictive modeling with loss

$\hat{s} = \Psi(r; K_{rx})$3

and at inference produces a future-frame embedding

$\hat{s} = \Psi(r; K_{rx})$4

The transmitted payload is then a predicted embedding sequence rather than raw frames, and a lightweight attentive probe at the vehicle decodes the received sequence into collision probabilities (Onsu et al., 23 Jan 2026).

A second family uses multi-modal semantic features and joint source–channel coding. In edge-intelligence-enabled autonomous driving, raw image, point cloud, and radar inputs are mapped to a common semantic space by a multi-branch DNN with inter-modal attention and contrastive alignment, $\hat{s} = \Psi(r; K_{rx})$5, then encoded for transmission over fading AWGN links, and reconstructed as $\hat{s} = \Psi(r; K_{rx})$6 at the edge server for downstream detection, classification, tracking, and segmentation (Feng et al., 2024). In the G-MSC framework, multi-modal fusion can be early, late, or intermediate through a BEV grid, with analog JSCC and digital transmission treated as alternative PHY instantiations depending on the V2X task (Lu et al., 2024).

A third family aligns semantic features with large multimodal or LLMs. In the LLaVA-based vehicle AI assistant, the vehicle extracts patch features $\hat{s} = \Psi(r; K_{rx})$7 using a compressed CLIP-ViT-L encoder, maps them to the language-model space through $\hat{s} = \Psi(r; K_{rx})$8, concatenates them with question embeddings, and performs VQA at the cloud server (Du et al., 5 May 2025). Semantic slicing, semantic matching via cosine similarity, fusion of objective and subjective attention, and patch-wise power allocation are used to prioritize image regions of greatest interest. In SemAgent, RSUs or vehicles derive compact feature vectors and semantic reports, semantically encode them as $\hat{s} = \Psi(r; K_{rx})$9, $x$0, or $x$1, transmit them over Rayleigh-fading plus AWGN channels, and combine the reconstructed features with historical trajectories in a prompt-driven LLM for future trajectory prediction (Zhu et al., 30 Nov 2025).

A fourth family uses symbolic or graph-like semantic abstractions. SEE-V2X begins with object detection and scene-graph generation at the RSU, followed by semantic prioritization and nonlinear transform source channel coding. The decoder reconstructs high-level object parameters such as class, position, and occlusion status, which can then be rendered as an AR “see-through” overlay for hidden pedestrians (Sun et al., 2 Sep 2025). Across these designs, semantic representation is task-specific: future latent tokens for collision prediction, BEV features for cooperative perception and prediction, scene graphs for hazard explanation, patch tokens for VQA, and trajectory-relevant summaries for forecasting.

3. Context modeling, relevance estimation, and message selection

The defining operation in task-oriented V2X is the estimation of what is relevant to whom, and under which context. In the context-aware paradigm of "The Search for Relevance" (Lusvarghi et al., 10 Aug 2025), object visibility is modeled probabilistically, with local detection probability expressed as a logistic function of distance. Relevance is then receiver-dependent: $x$2, where $x$3 denotes irrelevant or redundant information, and $x$4 denotes relevance under the current context. Transmission becomes a constrained top-$x$5 selection problem over sensed objects:

$x$6

The semantic packet then contains only the selected objects and minimal metadata such as object IDs, bounding boxes, and features (Lusvarghi et al., 10 Aug 2025).

A related formulation appears in the scalability analysis of joint semantic and task-oriented CPM generation. There, each vehicle first estimates redundancy by tracking objects already received in previous CPMs, then computes a binary relevance indicator

$x$7

for each intended receiver, aggregates it as $x$8, and solves a knapsack-like selection problem subject to CPM size limits. The result is a broadcast message containing only non-redundant objects with receiver impact (Lusvarghi et al., 8 Jun 2026). This gives a formal content-selection mechanism for dense broadcast settings rather than only for pairwise semantic links.

Pragmatic selection can also be spatial. In Select2Drive, the ego vehicle defines a Gaussian request map

$x$9

over the planned trajectory, while a supporting agent computes an alert signal from the difference between predicted remote semantics and the ego’s local estimate. Only pixels satisfying the thresholded product of request and alert are packed into the transmitted message. This is an approximate solution to a bandwidth-constrained priority maximization problem, and operationalizes the Area-of-Importance-based PragComm policy (Huang et al., 21 Jan 2025).

Decision tailoring can also be semantic rather than spatial. SEE-V2X maps driver tasks such as “avoid,” “slow down,” and “visualize hidden pedestrian” to decoder triggers, and renders or warns only about semantic elements whose utility exceeds a threshold $K_{tx}$0 (Sun et al., 2 Sep 2025). Taken together, these schemes indicate that relevance estimation may be object-centric, pixel-centric, path-centric, or explanation-centric, but the common principle is identical: the message is generated from anticipated receiver utility rather than from source completeness.

4. Communication models, PHY choices, and semantic-aware resource control

Semantic and task-oriented V2X modifies the communication stack above and below the semantic encoder. At the message level, payload sizes can be expressed explicitly. For V-JEPA collision prediction, raw video requires $K_{tx}$1 bytes, whereas a semantic embedding payload is $K_{tx}$2 bytes, with compression ratio $K_{tx}$3 (Onsu et al., 23 Jan 2026). This formulation makes the reduction in transmission burden a direct function of patch count, embedding dimension, and quantization precision.

At the PHY layer, the literature treats analog JSCC and digital transmission as complementary modes. In G-MSC, V2N is described as favoring digital transmission for long range and high security, V2I as digital if bandwidth allows or analog JSCC if ultra-low latency is required, and V2V/V2P as often favoring analog JSCC because of graceful degradation under low SNR and very low encoding delay (Lu et al., 2024). In the V-JEPA system, semantic payloads were evaluated over BPSK and QAM16 at uplink bandwidths of 10 MHz and 20 MHz; QAM16 at 20 MHz gave the lowest latency, while BPSK was more robust but higher latency (Onsu et al., 23 Jan 2026).

Semantic-aware resource management introduces new optimization criteria. In C-V2X platooning, semantic rate $K_{tx}$4, semantic similarity $K_{tx}$5, QoE, and Success Rate of Semantic Transmission (SRS) are used instead of purely bit-level rate and reliability. The joint problem allocates subchannels, powers, V2V versus V2I mode, and semantic symbol lengths so as to maximize platoon QoE while improving SRS, and is addressed by the SAMRA-MARL algorithm (Zhang et al., 2024). In 5G-V2X HetNets, semantic throughput is summarized through High-Speed Semantic Transmission Rate and High-Speed Semantic Spectrum Efficiency, with flexible duty-cycle coexistence between NR-U vehicular users and WiFi users. The SARADC-PPO framework jointly optimizes flexible duty cycle, resources, base stations, and semantic code granularity to maximize HSSE (Shao et al., 2024).

These resource-control formulations imply a shift in link adaptation. Instead of allocating power, channels, and code rate to maximize bit throughput for all payload bits uniformly, semantic-aware control allocates them to maximize semantic utility, semantic fidelity, or task success probability, sometimes while explicitly reducing the number of transmitted semantic symbols for less critical contexts (Zhang et al., 2024, Shao et al., 2024). A plausible implication is that semantic and task-oriented V2X requires cross-layer co-design: relevance estimation at the message-generation layer, semantic representation at the application layer, and PHY/MAC adaptation based on semantic criticality.

5. Representative systems and empirical performance

Collision prediction is a canonical demonstration. In the V-JEPA-based RSU-to-vehicle pipeline, the best post-processing method, binary road-masking, achieved $K_{tx}$6 F1 for collision prediction, compared with $K_{tx}$7 without post-processing, $K_{tx}$8 with heatmap highlighting, and $K_{tx}$9 with hybrid heatmap+mask. For early warning, the same system reported $K_{rx}$0 F1 at 12 frames before collision, $K_{rx}$1 at 8 frames, and $K_{rx}$2 at 4 frames. Communication volume dropped from approximately $K_{rx}$3 GB/clip for raw video to $K_{rx}$4 MB for FP32 embeddings, $K_{rx}$5 MB for FP16, and $K_{rx}$6 MB for INT8, corresponding to approximately $K_{rx}$7, $K_{rx}$8, and $K_{rx}$9 reduction, with QAM16 at 20 MHz and INT8 yielding latency on the order of $a = g(\hat{s})$0–$a = g(\hat{s})$1 ms per clip (Onsu et al., 23 Jan 2026).

Context-aware relevance filtering was evaluated both as a per-packet efficiency problem and as a network-scale broadcast problem. In unicast and broadcast simulations, the context-aware semantic scheme achieved roughly $a = g(\hat{s})$2 higher Semantic Efficiency than Baseline, IRC, and RM, delivered approximately $a = g(\hat{s})$3 more relevant information at $a = g(\hat{s})$4, reduced packet size by up to $a = g(\hat{s})$5, and maintained gains with statistical significance $a = g(\hat{s})$6 across more than $a = g(\hat{s})$7 random topologies (Lusvarghi et al., 10 Aug 2025). In the larger-scale CPM study, semantic and task-oriented content selection increased the number of supported vehicles by up to a $a = g(\hat{s})$8 factor under high-density conditions, decreased inter-reception time by up to $a = g(\hat{s})$9, and led to a twofold increase in the probability of successfully delivering all required relevant information to intended receivers (Lusvarghi et al., 8 Jun 2026).

Collaborative driving studies show similar task-level benefits. Select2Drive reported a $(\mathcal X,\mathcal Z)$0 improvement in offline perception tasks under limited bandwidth and a $(\mathcal X,\mathcal Z)$1 improvement under pose error conditions. In closed-loop driving, it delivered at most $(\mathcal X,\mathcal Z)$2 enhancement in driving score and $(\mathcal X,\mathcal Z)$3 enhancement in route completion rate, with the ablation showing that removing APC reduced driving score from $(\mathcal X,\mathcal Z)$4 to $(\mathcal X,\mathcal Z)$5 at $(\mathcal X,\mathcal Z)$6 MHz (Huang et al., 21 Jan 2025). This is a direct empirical instance of the “less is more” claim.

Large-model and generative systems extend the paradigm to question answering, BEV forecasting, and trajectory prediction. The LLaVA-SM framework maintained approximately $(\mathcal X,\mathcal Z)$7 VQA accuracy with $(\mathcal X,\mathcal Z)$8 tokens and $(\mathcal X,\mathcal Z)$9 T FLOPs, while reducing tokens by $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$0, FLOPs by $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$1, and response time by approximately $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$2 relative to LLaVA-1.6; under constrained channels, accuracy gains reached $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$3 at 12 dB and $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$4 at 10 dB, and fusion attention achieved $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$5 accuracy versus $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$6 for AVG-SemCOM (Du et al., 5 May 2025). In G-MSC, diffusion-model refinement improved BEV IoU by $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$7–$$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$8 points across three test scenes under low SNR of $$S:\mathcal X \rightarrow \mathcal Z,\; x \mapsto z=S(x)$$9–$\mathcal Z$0 dB, and increased semantic reliability from approximately $\mathcal Z$1 to approximately $\mathcal Z$2 at 3 dB (Lu et al., 2024). SemAgent reported up to approximately $\mathcal Z$3 improvement in prediction accuracy under low-SNR conditions and, in a V2I ablation at 20 dB, reduced FDE from $\mathcal Z$4 m for the LLM-only baseline to $\mathcal Z$5 m for the full model (Zhu et al., 30 Nov 2025).

Hazard communication and illustrative use cases further show the breadth of the paradigm. SEE-V2X simulations reported $\mathcal Z$6 m/s on a straight expressway at low density with RSU locations 1–2, $\mathcal Z$7 m/s at an intersection with moderate density and RSU near the center, and throughput up by approximately $\mathcal Z$8 in the latter case; when vehicles were already stabilized, marginal gains fell below $\mathcal Z$9 m/s (Sun et al., 2 Sep 2025). A broader survey article described semantic V2X use cases with more than $V_n$0 reduction in data volume for cooperative perception, $V_n$1 reduction in V2V load for platooning, uplink reduction from $V_n$2 Mbps to less than $V_n$3 Mbps for remote driving/task offloading, and semantic map updates with less than $V_n$4 bytes per update (Lyu et al., 2024). These results suggest that the empirical literature evaluates semantic and task-oriented V2X not only by compression ratio, but by its downstream effect on perception accuracy, control stability, throughput, early warning quality, and network scalability.

6. Limitations, trustworthiness, and open research directions

The current literature also documents substantive limitations. The V-JEPA collision-prediction system depends on digital-twin data, which may not capture all real-world lighting, weather, or sensor noise; its fixed patch size and embedding dimension may not be optimal for all scenarios; and it considers only a single RSU-to-vehicle flow without aggregation from multiple RSUs or vehicles. It also notes that strong YOLO detection is needed for post-processing and that mis-detections can hurt performance (Onsu et al., 23 Jan 2026). The relevance-search formulation assumes accurate context estimation, and the scalability study assumes perfect relevance estimation in the content-selection stage and an error-free V2X link in one of its evaluations, with explicit acknowledgment that real-world fading, packet loss, and hybrid ARQ or layered semantic coding remain open issues (Lusvarghi et al., 10 Aug 2025, Lusvarghi et al., 8 Jun 2026).

Security, privacy, and model consistency have become a separate line of inquiry. "Trustworthy Semantic Communication for Vehicular Networks" (Pan et al., 25 Sep 2025) proposes a three-layer VN-SemComNet architecture comprising semantic transmission, semantic encoding, and communication entity trust. Its semantic camouflage mechanism reported legitimate SSIM greater than $V_n$5 and an eavesdropper misleading rate of approximately $V_n$6. Its robust federated encoder–decoder training held SSIM at approximately $V_n$7 under $V_n$8 untargeted poisoning, compared with FedAvg SSIM of approximately $V_n$9 on MNIST, and its audit-game trust management increased global FL accuracy under $\hat{s} = \Psi(r; K_{rx})$00 adversaries to approximately $\hat{s} = \Psi(r; K_{rx})$01, compared with approximately $\hat{s} = \Psi(r; K_{rx})$02 without trust management. The same case study reported approximately $\hat{s} = \Psi(r; K_{rx})$03 latency reduction and approximately $\hat{s} = \Psi(r; K_{rx})$04 overhead reduction in a $\hat{s} = \Psi(r; K_{rx})$05-frame vehicular surveillance scenario (Pan et al., 25 Sep 2025).

Standardization and interoperability remain unresolved. Several works identify the need for common semantic ontologies, compact object-feature vocabularies, message formats, APIs, and protocol-stack support for relevance flags or compressed semantic content (Lyu et al., 2024, Lusvarghi et al., 10 Aug 2025, Lusvarghi et al., 8 Jun 2026). SEE-V2X further notes knowledge mismatch between RSU encoders and vehicle decoders and the need for continual model updates, while G-MSC and edge-intelligence studies emphasize edge-cloud orchestration, model partitioning, and adaptive semantic fidelity control under mobility and network load (Sun et al., 2 Sep 2025, Lu et al., 2024, Feng et al., 2024).

Open directions are therefore broadly convergent. The published agenda includes adaptive semantic compression by varying patch count or embedding dimension, multi-vehicle fusion and relay-based semantic reach, joint semantic V2X with ISAC, semantic-aware resource allocation, real-platform prototyping, semantic standardization, privacy-preserving knowledge-base updates, hybrid analog–digital semantic transmission, and extensions beyond collision prediction to traffic flow forecasting, pedestrian intent prediction, cooperative path planning, and trajectory prediction (Onsu et al., 23 Jan 2026, Lusvarghi et al., 10 Aug 2025, Lyu et al., 2024, Lu et al., 2024, Zhu et al., 30 Nov 2025). Taken together, these directions indicate that semantic and task-oriented V2X is evolving from an application-specific compression technique into a broader systems discipline spanning representation learning, context reasoning, cross-layer scheduling, trustworthy model sharing, and scalable vehicular network design.