X-to-Maritime: Adapting Methods for Marine Systems
- X-to-Maritime is a cross-domain program that retools methods, architectures, and datasets from non-maritime fields to suit maritime navigation, robotics, and decision support.
- It demonstrates significant gains, such as improved performance of maritime-specific models after fine-tuning and enhanced sensor fusion across various detection technologies.
- The approach integrates embodied AI, advanced mesh networking, simulation-based testbeds, and phased certification to meet the complex demands of maritime environments.
X-to-Maritime, as reflected in recent literature, denotes a recurring cross-domain adaptation pattern in which methods, architectures, datasets, infrastructures, and assurance practices developed outside the maritime domain are reworked for maritime navigation, communications, robotics, perception, logistics, and decision support. The pattern appears explicitly in several forms: embodied large models for intelligent vessels, multi-modality-conditioned maritime generation, Wi‑Fi 6 maritime mesh networking, quantum just-in-time navigation, historical knowledge graphs from AIS, and Simulation-to-Maritime, Lab-to-Maritime, and Space-to-Maritime transfer in controlled testbeds. This suggests that X-to-Maritime is not a single method but a broad technical program for translating non-maritime capabilities into environments characterized by multi-source heterogeneous data, harsh weather, open line-of-sight propagation, long-range sensing, regulatory constraints, and safety-critical human oversight (Wang et al., 2024, Torroba et al., 26 Feb 2026, Guo et al., 25 Sep 2025, Li et al., 19 Feb 2025, Imrecke et al., 2021).
1. Conceptual scope and representative forms
Across the cited work, X-to-Maritime is used to transfer multiple kinds of “X” into maritime settings: foundation models and embodied agents into intelligent vessels; generic visual trackers into maritime visual object tracking; diffusion-based scene generation into maritime object detection; Wi‑Fi 6 mesh networking into offshore connectivity; knowledge graphs and Markovian mobility models into AIS-driven ETA and behavior analysis; and space-robotics and simulation infrastructures into maritime experimentation. This suggests a unifying formulation: maritime adaptation requires not only task transfer but also redefinition of sensing, physical constraints, data curation, and evaluation criteria.
| Transferred “X” | Maritime instantiation | Representative paper |
|---|---|---|
| Embodied large models | KUNPENG intelligent vessel “AI brain” | (Wang et al., 2024) |
| Multi-modality-conditioned diffusion | Neptune‑X / X-to-Maritime scene generation | (Guo et al., 25 Sep 2025) |
| Generic VOT trackers | MVTD maritime fine-tuning benchmark | (Bakht et al., 3 Jun 2025) |
| Wi‑Fi 6 mesh networking | AX‑MMN in the Red Sea | (Li et al., 19 Feb 2025) |
| Knowledge graphs from AIS | Global ETA prediction | (Dimitriou, 18 May 2026) |
| Quantum optimization | Maritime just-in-time navigation | (Imrecke et al., 2021) |
| Space-robotics and sim-to-real | Marinarium | (Torroba et al., 26 Feb 2026) |
A central theme is that maritime deployment rarely succeeds through zero-shot reuse of terrestrial or generic systems. Neptune‑X begins from latent diffusion but introduces Bidirectional Object-Water Attention and Attribute-correlated Active Sampling; MVTD shows that 14 recent SOTA tracking algorithms degrade substantially on maritime video and improve only after maritime fine-tuning; AX‑MMN repurposes IEEE 802.11ax but depends on coastal base stations, relay stations on uninhabited islands, moored buoys, solar power, and mesh topology; and the MASS certification literature imports aerospace- and automotive-style phased assurance into a maritime ladder of simulation, HIL, small craft trials, medium ship trials, and IMO certification (Guo et al., 25 Sep 2025, Bakht et al., 3 Jun 2025, Li et al., 19 Feb 2025, Arnaoot, 2 Apr 2025).
2. Architectural transfer: from embodied AI to vessel-scale autonomy
A prominent architectural instance of X-to-Maritime is KUNPENG, described as “the first-ever embodied large model for intelligent maritime in the smart ocean construction.” It treats the vessel as an embodied AI system with an integrated brain and decomposes the platform into six interacting systems: Perception System, Cognition System, Decision System, Behavior System, Power System, and Safety & Emergency System. Its backbone combines LLaMA‑3 with Perceiver/Perceiver‑IO, and its control loop is explicitly Perception → Cognition → Decision → Behavior, with additional power scheduling and safety/emergency handling (Wang et al., 2024).
The Perception System fuses historical navigation data, distributed sensor data, vessel status, and external hydrometeorological and satellite data into a unified latent representation. The Cognition System builds “Embodied Route Environment Cognition,” includes hydrometeorological forecasting, and implements “Theory of Mind (ToM) for vessel interaction cognition,” treating other vessels as agents with their own intents and trajectories. The Decision System uses imitation learning for decision, autonomous berthing and unberthing, intelligent route navigation, and autonomous navigation obstacle avoidance, while the Behavior System implements a “Dual‑Pathway Fast‑Slow Behavioral Model” in which a fast path supports real-time collision avoidance and emergency maneuvers and a slow path supports long-term optimization (Wang et al., 2024).
This architecture also shows how maritime transfer forces broader system closure than many land-based AI deployments. KUNPENG integrates embodied energy consumption management, intelligent fault diagnosis, liability determination, compliance checking, and emergency planning. A plausible implication is that X-to-Maritime architectures must absorb hydrometeorology, legal reasoning, and power scheduling into the same control graph rather than treating them as external services. The paper’s own framing is correspondingly broad: the model is intended for smart vessels, route optimization, safe navigation, and intelligent maritime operations, rather than for a single perception or control subtask (Wang et al., 2024).
3. Data, representation, and benchmark transfer
A large part of X-to-Maritime is data-centric. In sensing and perception, MOANA presents “a comprehensive maritime sensor dataset featuring multi-range detection capabilities” by integrating short-range LiDAR, medium-range W-band radar, and long-range X-band radar into a unified framework, with object labels derived from radar and stereo camera images and seven sequences collected from diverse regions. The dataset is intended for place recognition, odometry estimation, SLAM, object detection, and dynamic object elimination within maritime environments, thereby importing automotive and robotics-style dataset design into maritime sensing (Jang et al., 2024).
The same pattern appears in visual tracking. MVTD comprises 182 high-resolution video sequences totaling approximately 150,000 frames across four representative object classes—boat, ship, sailboat, and unmanned surface vehicle—and evaluates 14 recent SOTA tracking algorithms. The generic-to-maritime gap is explicit: pretrained trackers such as HIPTrack, MCITrack, SimTrack, and GRM degrade on MVTD, while maritime fine-tuning yields large gains; for example, SimTrack improves from AUC 67.01 to 75.14, HIPTrack from 68.65 to 74.43, and MCITrack from 67.78 to 72.49. The benchmark therefore operationalizes maritime adaptation as measurable domain transfer rather than as a purely conceptual claim (Bakht et al., 3 Jun 2025).
Neptune‑X generalizes the data-centric strategy further by making generation itself maritime-specific. Its X-to-Maritime module is a multi-modality-conditioned generative model built on Stable Diffusion and trained on the Maritime Generation Dataset, which contains 11,900 images with bounding boxes, object categories, water surface masks, image captions, and water surface descriptions. The model conditions on global text, object-level labels and layouts, and water masks, and introduces the Bidirectional Object-Water Attention module to capture boundary interactions between objects and their aquatic surroundings. Empirically, the generator achieves FID 18.05, CAS 79.34, and YOLO Score 17.08 / 39.14 / 13.52, and when its selected synthetic data are used to augment YOLOv10, mAP improves from 39.99 / 61.13 to 43.62 / 65.50 (Guo et al., 25 Sep 2025).
Maritime transfer also extends to spatio-temporal prediction and graph representation. FLP‑XR adapts extreme-scale tabular ML rather than sequence-heavy deep models: it formulates future location prediction as
uses AIS-derived features with XGBoost, and reports 2–3 orders of magnitude faster training and inference than the current state of the art. On the Brest dataset, training is 69 s for FLP‑XR versus 35,909 s for Nautilus, with inference 0.56 s versus 1345 s per AIS message, while maintaining competitive or superior Haversine error in many horizons (Theodoropoulos et al., 10 Mar 2025). At the level of structured maritime analytics, “Historical Knowledge Graphs for Global Maritime Estimated Time of Arrival” constructs a graph with 5,433 geohash-3 nodes and 12,334 edges using only AIS data, achieving median RMSE 22.75 min at segment level and 30.90 min at trajectory level on a temporally held-out set, with 69.13% of trajectories within 20% of actual arrival time; “Learning Spatio-Temporal Vessel Behavior using AIS Trajectory Data and Markovian Models in the Gulf of St. Lawrence” discretizes the ocean into about 8,687 H3 cells and finds vessel-specific mobility signatures and significant but transient pandemic-induced deviations (Dimitriou, 18 May 2026, Spadon et al., 22 May 2025).
4. Communications and network infrastructures
In communications, X-to-Maritime frequently means extending terrestrial, aerial, or integrated wireless architectures into sea-space topologies. AX‑MMN is an especially direct case: it adapts IEEE 802.11ax into a maritime mesh network built from coastal terrestrial base stations, relay stations on uninhabited islands, moored buoys, and fishermen’s mobile devices. In a Red Sea deployment covering about 1 km² with eight nodes, the system achieves practical field results of 20–90 Mbps to standard phones, and the entire 8-node testbed cost is approximately \$1,489, while satellite phones are described as providing 2.2–9.6 kbps at far higher device and service costs. The network operated autonomously for two months with solar power on all nodes except the terrestrial gateway (Li et al., 19 Feb 2025).
A different line of work transfers optimization-based aerial networking into maritime hybrid systems. “Maritime Coverage Enhancement Using UAVs Coordinated with Hybrid Satellite-Terrestrial Networks” studies a UAV that accompanies a ship and jointly optimizes trajectory and in-flight transmit power using only location-dependent large-scale CSI, subject to tolerable interference, backhaul, kinematic, and communication-energy constraints. The problem is formulated as a max–min ergodic-rate design and solved by decomposition, successive convex optimization, and bisection. The results show distinct energy-limited and interference-limited regimes and demonstrate that UAV fits well with existing satellite and terrestrial systems under the proposed framework (Li et al., 2019).
More recent work expands this to secure low-altitude and space-air-ground-sea systems. “Secure Low-altitude Maritime Communications via Intelligent Jamming” formulates a secure and energy-efficient maritime communication multi-objective optimization problem as a POMDP and proposes SAC‑CVAE, a soft actor-critic with conditional variational autoencoder using advantage-conditioned latent representations and LSTM-based eavesdropper prediction. The framework jointly controls Alice and Bob UAV trajectories and powers to maximize secrecy rate while minimizing energy, and simulation results show that intelligent jamming yields positive secrecy where non-jamming cases have secrecy rate essentially 0 or negative (Huang et al., 10 Nov 2025). “Unified Design of Space-Air-Ground-Sea Integrated Maritime Communications” divides maritime space into coastal, offshore, middle-sea, and open-sea regions served respectively by TBS, USV, UAV, and satellite, and designs a joint beamforming and trajectory optimization algorithm to maximize the minimum transmission rate. The proposed algorithm reaches Jain’s index close to 1, and the reported trajectories show that USV and UAV mobility is used not only for distance reduction but also for interference shaping across layers (Zhou et al., 13 Aug 2025).
5. Simulation, testbeds, human oversight, and certification
X-to-Maritime also concerns where and how maritime systems are validated. Marinarium is a modular and stand-alone underwater research facility designed to bridge simulation, laboratory validation, and field conditions, and the paper explicitly frames it as enabling Simulation-to-Maritime, Lab-to-Maritime, Space-to-Maritime, and aerial/surface/underwater-to-maritime transfer. The facility combines a water basin, above-water robotic volume, retractable roof, integrated motion capture, and a digital twin in SMaRCSim. The testbed supports learning-based system identification, heterogeneous rendezvous missions, and spacecraft-surrogate validation; one concrete infrastructure result is that adding sea gravel improved Delphis Succorfish acoustic modem packet reception from 0% to 70% (Torroba et al., 26 Feb 2026).
The same article demonstrates how controlled maritime environments can serve as intermediate transfer layers rather than final deployment targets. For BlueROV2, a Koopman EDMDc–RBF model trained on tank data reaches 1-step / 10-step / 100-step RMSE of 0.0629 / 0.0831 / 0.1859, outperforming both a double-integrator baseline and a Fossen BlueROV2 physics model. For the SAM AUV, a learned residual model reduces positional endpoint RMSE at horizon from 0.72 m to 0.32 m. This suggests that in X-to-Maritime, sim-to-real transfer is often mediated by a controlled maritime testbed rather than by direct field deployment (Torroba et al., 26 Feb 2026).
Human-centered adaptation is equally visible in explainability and certification. “From Sea to System: Exploring User-Centered Explainable AI for Maritime Decision Support” proposes a domain-specific survey framework to measure trust, usability, and explainability for maritime professionals using radar-based navigational scenarios, a maritime assistant, and key explanatory features. The paper does not yet report empirical results, but it frames maritime XAI as support for cognitive trust, affective trust, situation awareness, and bridge-team decision making under COLREG-constrained collision avoidance (Jirak et al., 18 Sep 2025). In parallel, “Navigating the Uncharted Waters” imports phased certification logic from other safety-critical domains into MASS and defines a six-phase ladder: path plan algorithm development, HIL controller testing, initial trials with small unmanned rubber boat, small craft trials ( m), medium ship trials (15–100 m), and IMO certification trials. The final IMO phase requires 50,000 nautical miles, zero collisions, navigation error, system failure rate, and 99.999% availability, while the paper’s timing analysis concludes that the update rate must be at least 10 Hz for sufficient safety margin in its collision scenario (Arnaoot, 2 Apr 2025).
6. Limitations, controversies, and future directions
A recurring limitation is that maritime transfer is often data-limited, geographically biased, or only partially benchmarked. Neptune‑X identifies scarcity of annotated maritime data and poor generalization across object category, viewpoint, location, and imaging environment; MVTD shows substantial performance degradation of generic trackers before maritime fine-tuning; and the historical ETA graph, although global and AIS-only, cannot account for severe weather events, port closures, temporary traffic separation schemes, or canal disruptions because it uses only historical movement patterns (Guo et al., 25 Sep 2025, Bakht et al., 3 Jun 2025, Dimitriou, 18 May 2026). This suggests that X-to-Maritime cannot be reduced to architecture transfer alone; representation coverage and attribute balance remain first-order constraints.
Another limitation is that several influential papers are still more architectural or positioning frameworks than fully benchmarked operational systems. KUNPENG is presented as a blueprint for embodied maritime foundation models, but the available text explicitly states that it “does not give explicit numerical results, metrics, or tables” and should currently be understood more as a conceptual and architectural contribution; the user-centered XAI paper is likewise a design and survey framework without reported empirical findings; and the MASS certification paper proposes staged metrics and thresholds but does not constitute a regulatory code in itself (Wang et al., 2024, Jirak et al., 18 Sep 2025, Arnaoot, 2 Apr 2025). A common misconception is therefore that X-to-Maritime necessarily means mature deployment; in many cases it instead denotes an intermediate research program spanning dataset construction, simulation, and staged validation.
Future directions in the literature are correspondingly multi-layered. Communications papers call for robust design under uncertain ship locations and channel statistics, multi-UAV coordination, integrated satellite–UAV–terrestrial routing, and energy-aware propulsion-and-communication optimization (Li et al., 2019). Secure maritime LAWNs introduce generative RL but still face practical issues of regulation, hardware budgets, and adaptive adversaries (Huang et al., 10 Nov 2025). Space-air-ground-sea integration highlights the need for robust CSI handling, energy constraints, and scalable distributed algorithms (Zhou et al., 13 Aug 2025). In perception and data, Neptune‑X points toward richer modalities and continuous attributes, while MVTD and MOANA indicate that maritime benchmarks for tracking, multi-radar odometry, and navigation remain foundational infrastructure rather than solved endpoints (Guo et al., 25 Sep 2025, Bakht et al., 3 Jun 2025, Jang et al., 2024). Taken together, the literature suggests that X-to-Maritime is evolving from straightforward domain adaptation into a broader synthesis of data-centric generation, physically grounded sensing and networking, embodied autonomy, human-centered oversight, and certification-aware system engineering.