Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems

arXiv:2607.13880 · cs.LG · Submitted 2026-07-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems".

Jane: Here's the gist of the paper. The gist The proposed Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL) framework leverages practical information to characterize the utility of information for cooperative tasks and learns distributed cooperative policies under realistic communication and covert constraints.

Tom: Problem Context Autonomous underwater vehicles (AUVs) often cannot rely on active sonar to continuously obtain complete information, since active sensing and frequent communications increase the risk of exposure <ref:2607.13880#pg2> AUVs primarily rely on passive observation, an approach that yields incomplete local perception and limited task efficiency <ref:2607.13880#pg2>.

Jane: Underwater spaces inherently suffer from severe propagation attenuation, long propagation delays, narrow bandwidth, and strong dynamics, which pose significant challenges to information exchange and cooperative decision-making <ref:2607.13880#pg2>.

Tom: Active sonar's high-power acoustic radiation can be directly detected by distant devices, making it difficult to satisfy the survivability requirements of covert missions <ref:2607.13880#pg3>.

Paper summary: Tom: So we're looking at this paper now titled "Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems". Jane, what's the core idea they’re pushing here?

Jane: Well, it starts by pointing out that autonomous underwater vehicles, or AUVs, can’t just use active sonar all the time because sending those signals and communicating constantly puts them at risk of being detected.

Tom: Right. So instead of relying on that active sensing to get everything they need, the paper says AUVs mostly rely on passive observation.

Lu: And that passive observation is a problem because it means they don't get complete local perception, which limits how well they can actually finish their tasks >-ref:two thousand six hundred seven point one three eight eight zero#pg2>

Jane: Exactly, and then you bring in communication to help fix that incomplete picture, but the paper notes that using underwater acoustic communication has its own set of serious headaches.

Tom: Like long delays, interference from other things underwater, low reliability for those messages, and the actual risk of being exposed while they're talking >-ref:two thousand six hundred seven point one three eight eight zero#pg2>

Lu: And what makes this paper different is that existing multi-agent reinforcement learning studies often treat communication like it’s a perfect, ideal information flow, whereas traditional communication optimization only looks at how reliable the physical link itself is >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Jane: So what they claim here in "Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems" is that both of those old ways are actually insufficient because they don't account for how much the actual information helps the mission, considering all these realistic physical and covert constraints >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Tom: It proposes this SVR-MARL framework, which stands for Sensed Information Value Realization Multi-Agent Reinforcement Learning. Basically, they leverage the practical information they sense to figure out the real utility of that information for the team's tasks >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Lu: They break down each AUV’s decision into two parts: a communication phase and a task-execution phase, with an action being defined as a i = a comm i, a task i >-ref:two thousand six hundred seven point one three eight eight zero#pg6>

Jane: And they introduce this concept of the "value of sensed information," which they use to figure out the potential collaborative benefit that local observations might give to other AUVs >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Tom: This value is then shaped by those realistic communication and covert constraints, showing how it actually contributes to the collaboration >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Lu: The framework has a few key pieces: they model the task and how information exchange helps collaboration, they have coupling mechanisms where covertness links the communication and task phases because sending a message changes the risk of being exposed, and then there's this Value Realization Mechanism for that sensed information >-ref:two thousand six hundred seven point one three eight eight zero#pg7>

Jane: They treat the environment as a physics-grounded covert communication environment, explicitly considering things like propagation delays, concurrent interference, and covertness in the acoustic communication part >-ref:two thousand six hundred seven point one three eight eight zero#pg7>

Tom: And they design a specific reward for the communication phase called r comm, which is meant to characterize the practical information value of a message by balancing its potential benefit against those real costs >-ref:two thousand six hundred seven point one three eight eight zero#pg7>

Paper summary: Lu: The case study they use is covert multi-AUV cooperative localization and tracking, and the results show that this framework can improve task efficiency while cutting down on unnecessary communication and exposure risks >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Jane: The results are quite solid here; they showed a task efficiency of approximately one point two times ten-two which they said outperformed IC3Net by about twenty percent and was the best among all the practical communication schemes tested >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Tom: And they have a neat finding where it adapts transmit power based on task needs; so it increases communication when target uncertainty is high, and then reduces unnecessary transmissions once the target is well localized >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Lu: That suggests that this value-driven communication scheduling can really improve target localization and task completion efficiency even when you only have bearing-only passive observations >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Jane: So, what does this mean for a listener who just wants to know what it does? It means for covert missions, they can use this SVR-MARL framework to make sure the AUVs aren't just talking around in circles but are actually using their limited information wisely >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Tom: And it shows that you don't need perfect, ideal communication to get good results; you can achieve better task performance by making smart decisions about when and how much to transmit >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Jane: It really moves the focus from just "can we send a message?" to "is sending *this specific* message right now worth the risk and delay?" which is a big shift in how we think about these systems >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Lu: Moving on to the conclusion of this paper, they are focusing on two main future directions. First, they want to add a teammate intention prediction module so each AUV can predict what other AUVs plan to do next >-ref:two thousand six hundred seven point one three eight eight zero#pg7>

Tom: And second, there's work on secure agent authentication and trustworthy communication mechanisms using things like acoustic fingerprints and channel characteristics >-ref:two thousand six hundred seven point one three eight eight zero#pg7>

Jane: So, to wrap up the whole thing on "Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems," it’s about taking the messy reality of underwater communication—the delays, the interference, and the covert risk—and building a way for AUVs to use their sense of information intelligently >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Lu: The authors are clearly arguing that by modeling this as a value realization problem under those constraints, they can learn distributed cooperative policies that are actually physically realizable >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Tom: It’s about using practical information to characterize the utility for tasks and then learning those distributed policies while respecting all the communication and covert limitations >-ref:two thousand six hundred seven point one three eight eight zero#pg3>

Jane: So, it suggests that for future systems, we need to stop treating communication as a perfect pipe and start treating it as a constrained resource whose value depends on the actual information it carries >-ref:two thousand six hundred seven point one three eight eight zero#pg5>

Conclusion: Tom: So, we’re wrapping up our look at "Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems." This paper is about how autonomous underwater vehicles can actually do their jobs better when they have to stay hidden.

Jane: Right, so the core idea is that because active sonar and constant talking are too risky, these AUVs have to rely on listening passively for information, which usually leaves them blind locally.

Lu: And the paper tackles this by saying we need a way to figure out what value that sensed information actually has for the team's task, not just how much data is there.

Meng: From a deployment standpoint, that means if you deploy these vehicles in a covert setting, they shouldn't waste time sending messages that don't help the mission at all.

Tom: Exactly. It’s not just about communication links anymore; it’s about valuing the actual content of the message under real-world underwater constraints like delay and interference.

Jane: The authors propose this SVR-MARL framework to learn these distributed policies, balancing what they sense with what they need to do for their mission.

Lu: The key mechanism is this Value Realization part, where the system estimates the potential benefit of local observations before deciding whether to transmit anything.

Tom: And they show that in a real test case—covert tracking and localization—this approach gave them a task efficiency about one point two times ten to the negative two.

Jane: That number is significant because it shows this value-driven scheduling actually helps them do the job faster than just sending everything they can.

Meng: But I wonder, how much extra computation does this SVR-MARL framework add to the vehicle’s onboard processing unit when you have to calculate all those values and constraints?

Lu: That’s a fair question, Meng. The complexity comes from coupling the communication and task phases together in that closed-loop POMG structure they describe.

Tom: Yeah, it couples them, so the decision to talk affects the risk of exposure, and that changes what information you can get next.

Jane: So for someone listening who isn't deep in reinforcement learning, it means this system learns to be smart about when to whisper versus when to stay silent underwater.

Lu: It shifts the focus from just link reliability—which is what traditional communication optimization cares about—to the actual utility of the sensed data for collaboration.

Meng: What they didn't really cover, and I’m curious if it matters for production, is how robust this whole model is when you introduce a completely new type of interference pattern we haven't seen before.

Tom: That’s a huge question about generalization. The paper focuses on the known constraints like delay and interference, but scaling it to totally unknown environments is where things get tricky.

Jane: So, the big implication here is that for any future covert mission, if you can figure out what information matters most locally, you can design a system that uses it efficiently without constantly broadcasting.

Lu: And they point toward future work involving teammate intention prediction and secure authentication based on acoustic fingerprints to make those learned policies even more reliable in the field.

Xueyao Zhang, Chenyang Yan, Bo Yang, Xuelin Cao, Zhiwen Yu, Bin Guo, George C. Alexandropoulos, Merouane Debbah, Á Chau Yuen

Northwestern Polytechnical University School of Computer Science Department of Informatics and Telecommunications National and Kapodistrian University of Athens KU 6G Research Center Khalifa University CentraleSupelec Nanyang Technological University

cs.LG

Submitted: 2026-07-15

Updated: 2026-10-19

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: The gist The proposed Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL) framework leverages practical information to characterize the utility of information for

Key concepts

Sensed-information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL): This is a novel learning framework designed for multi-AUV systems. It focuses on calculating the actual collaborative benefit (value) derived from local observations (sensed information) and then learns policies that use this value to decide when and how much to communicate, while respecting real-world constraints.
Communication vs. Task Optimization: Existing methods separate communication optimization (focusing on link reliability) from task optimization (focusing on message content). This paper integrates them by designing a reward system that balances the practical benefit of a message with the actual costs of covert transmission and realistic acoustic channel conditions.
Covertness Coupling: This concept links communication decisions directly to mission risk. The decision to transmit information affects how exposed an AUV is, and conversely, the physical action taken during a task can influence this exposure through geometric factors. This ensures the learned policies account for survival requirements.
Value of Sensed Information: This is a measure used in the framework to quantify how much local sensing data (like bearing information) will actually help other AUVs achieve their cooperative goals. By estimating this value, the system learns to prioritize sharing only when the potential collaborative gain outweighs the risks associated with communication.

Terminology

Reinforcement Learning — 2× in this paper · explained in 4 episodes
Multi-Agent Reinforcement Learning (MARL) — 1× in this paper · explained in 4 episodes

Summary

The gist The proposed Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL) framework leverages practical information to characterize the utility of information for cooperative tasks and learns distributed cooperative policies under realistic communication and covert constraints.

Problem Context

Autonomous underwater vehicles (AUVs) often cannot rely on active sonar to continuously obtain complete information, since active sensing and frequent communications increase the risk of exposure <ref:2607.13880#pg2> AUVs primarily rely on passive observation, an approach that yields incomplete local perception and limited task efficiency <ref:2607.13880#pg2>. Underwater spaces inherently suffer from severe propagation attenuation, long propagation delays, narrow bandwidth, and strong dynamics, which pose significant challenges to information exchange and cooperative decision-making <ref:2607.13880#pg2>. Active sonar's high-power acoustic radiation can be directly detected by distant devices, making it difficult to satisfy the survivability requirements of covert missions <ref:2607.13880#pg3>. Passive observations usually provide only limited local bearing information, making it difficult for an individual AUV to accurately recover the target state and potentially leading to accumulated perception uncertainty and unstable localization <ref:2607.13880#pg2>.

Limitations of Existing Approaches

Existing communication-oriented multi-agent reinforcement learning (MARL) studies often model communication as an ideal information flow, whereas traditional communication optimization primarily focuses on link-level performance <ref:2607.13880#pg3>. These existing methods characterize “information interaction” in cooperative communication and “link reliability” in physical communication, respectively, but they remain insufficient for the more practical underwater multi-AUV covert collaboration scenario considered in this work <ref:2607.13880#pg3>. Specifically, MARL-based message passing typically lacks realistic underwater acoustic modeling, including propagation delays, interference, decoding reliability, and covert constraints <ref:2607.13880#pg4>. Conversely, traditional communication optimization focuses on the physical transmission process and mainly aims to improve communication performance itself without considering the task significance of the message content <ref:2607.13880#pg4>.

SVR-MARL Framework Design

The SVR-MARL framework introduces the concept of “value of sensed information” to characterize the potential collaborative benefit inherent in local observations <ref:2607.13880#pg5>. This value is reshaped by realistic communication and covert constraints, reflecting its contribution to collaboration <ref:2607.13880#pg5>. The framework decomposes each AUV’s policy into a communication phase and a task-execution phase, with the corresponding action expressed as ai = acomm i, atask i <ref:2607.13880#pg6>.

The framework involves several key components:

Task Modeling and Action Decomposition: The common structure is divided into how to exchange information that benefits collaboration and how to perform task-oriented physical actions <ref:2607.13880#pg6>.
Coupling Mechanisms: Covertness couples the communication and task phases, where the transmission decision affects exposure risk, and the task action influences covertness through geometry <ref:2607.13880#pg6>. A closed-loop coupling in POMG is established where communication outcomes change the information available to affect subsequent task observations <ref:2607.13880#pg6>.
Value Realization Mechanism of Sensed Information: This module estimates the potential collaborative benefit that current-sensing information may provide to other AUVs, which is then used as input to the communication policy network <ref:2607.13880#pg7>.
Physics-Grounded Covert Communication Environment: The environment treats communication as the key realization layer connecting ”value of sensed information” and ”practical information value,” explicitly considering delay, concurrent interference, and covertness in underwater acoustic communication <ref:2607.13880#pg7>.
Practical Information Value-Guided Communication Phase: The communication reward rcomm is designed to characterize the practical information value of a message by balancing its counterfactual benefit with realistic communication and covert costs <ref:2607.13880#pg7>.

Case Study Results

In a case study of covert multi-AUV cooperative localization and tracking, the proposed framework was validated against three comparative schemes: True Obs scheme, Full comm, and No comm. The results show that the proposed framework achieves a task efficiency of approximately 1.2×10−2, outperforming IC3Net by about 20% and achieving the best performance among all practical communication schemes. Furthermore, the proposed framework adapts transmit power according to task requirements, increasing communication investment when target uncertainty is high and reducing unnecessary transmissions once the target is well localized. This indicates that value-driven communication scheduling can effectively improve target localization and task completion efficiency under bearing-only passive observations.

Future Directions

Future work will focus on two main areas:

Teammate Intention Prediction for Distributed Coordination: A teammate-intention prediction module will be introduced to enable each AUV to recursively estimate other AUVs’ communication tendencies and action intentions over the next few steps.
Secure Agent Authentication and Trustworthy Communication: Future work can investigate secure authentication and trustworthy message evaluation mechanisms based on acoustic fingerprints, channel characteristics, motion behaviors, and historical communication patterns.

--- Page 1 ---

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 1

Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems

Xueyao Zhang, Chenyang Yan, Bo Yang, Xuelin Cao, Zhiwen Yu, Bin Guo, George C. Alexandropoulos, Merouane Debbah, and Chau Yuen ´

Abstract—In underwater covert cooperative missions <ref:2607.13880#pg2> autonomous underwater vehicles (AUVs) often cannot rely on active sonar to continuously obtain complete information, since active sensing and frequent communications increase the risk of exposure <ref:2607.13880#pg2>. As a result, AUVs primarily rely on passive observation, an approach that yields incomplete local perception and limited task efficiency <ref:2607.13880#pg2>. Although underwater acoustic communications can mitigate this limitation through information sharing, they are simultaneously constrained by long delays, severe interference, low reliability, and the risk of covert exposure <ref:2607.13880#pg2>. Existing communications-oriented multi-agent reinforcement learning (MARL) studies often model communication as an ideal information flow, whereas traditional communication optimization primarily focuses on link-level performance <ref:2607.13880#pg3>. However, both are insufficient to characterize the actual contribution of perceptual information to cooperative tasks under realistic conditions of covert physical communications <ref:2607.13880#pg3>. This paper proposes a Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL) framework that leverages practical information to characterize the utility of information for cooperative tasks and learns distributed cooperative policies under realistic communication and covert constraints <ref:2607.13880#pg3>. Through a case study of covert multi-AUV cooperative localization and tracking, the potential of the proposed framework to improve collaborative task efficiency while reducing unnecessary communication and exposure risks is demonstrated <ref:2607.13880#pg3>. Index Terms—Autonomous underwater vehicles, multi-agent reinforcement learning, covert transmission <ref:2607.13880#pg2>.

--- Page 2 ---

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 2

Single AUV Local uncertainty Collaboration demand Communication helps

Collaboration demand Our scenario AUV i AUV N Local sensing Message Covert risk Eavesdropper Low efficiency High efficiency Ideal Communication Communication-centric Optimization The Problem MARL Collaboration Communication-centric Optimization Policy network O1 Oi On End-to-end Optimization Inputs Outputs Network Information Resource Allocation Qos Constraint Reliable Links High Throughput Low Delay Problem Type: MINLP, Multi-objective Methods: Alternating optimization, SCA. MARL-based collaboration focuses on task success, but commonly assumes ideal communication and ignores realistic transmission constraints <ref:2607.13880#pg3>. Communication optimization focuses on ensuring reliable transmissions, regardless of whether the transmitted information is critical for collaboration <ref:2607.13880#pg3>.

--- Page 3 ---

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 3

TABLE I COMPARISON OF DIFFERENT COMMUNICATION APPROACHES. Block Representative Works Realistic Channel Constraints Covert Transmission Task-oriented Value MARL Collaboration with ideal communication Sukhbaatar et al. [6] × × Singh et al. [4] × × Das et al., Niu et al. [12] × × Our work SVR-MARL considered.

Improvements for AI systems

Value of Sensed Information Estimation allows AUVs to determine if a local observation truly contributes to the team task before transmission, optimizing for practical information value. This enables agents to prioritize messages based on actual utility for team collaboration rather than just importance or link availability.
Closed-loop coupling in the POMG structure ensures that communication and task phases are not independent; instead, a change in acoustic link outcome (like delay or decoding failure) directly affects the next task observation, creating a continuous feedback loop for adaptive decision-making.
The framework allows AUVs to learn when to communicate and which information is worth transmitting under realistic constraints by jointly considering information value, link availability, and covert requirements during communication decision-making. This prevents redundant communication and reduces unnecessary exposure risks.
Task decomposition into a communication phase and a task-execution phase allows AUVs to independently learn when to decide on transmission parameters (like transmit power) versus executing physical actions (like velocity or maneuver). This modularity simplifies policy learning by separating information selection from motion planning.
The system can be improved with a teammate intention prediction module that enables each AUV to recursively estimate other AUVs’ communication tendencies and action intentions over the next few steps. This prediction helps agents better infer the team's collaborative state, leading to more effective future communication and coordinated actions.
Secure agent authentication and trustworthy message evaluation mechanisms can be integrated based on acoustic fingerprints, channel characteristics, motion behaviors, and historical communication patterns. This would allow the system to distinguish legitimate nodes from malicious ones before message fusion, enhancing robustness against adversarial attacks.

Sources

Cooperative Target Detection with AUVs: A Dual-Timescale Hierarchical MARDL Approach