CoralGuide: Tethered Robotics & Reef Analysis

• CoralGuide is a framework that integrates tethered marine robotics for precise path planning with advanced coral reef mapping and analysis.
• It employs techniques like A* search, Bézier interpolation, and catenary modeling to ensure safe, obstacle-free multi-robot trajectories.
• The ecosystem extends to multimodal coral assessment through georeferenced mosaicing, semantic segmentation, and restoration strategies.

Searching arXiv for CoralGuide and closely related coral-reef monitoring papers to ground the article. CoralGuide denotes, most explicitly, a framework for path planning and trajectory optimization in tethered marine multi-robot systems composed of an autonomous surface vehicle (ASV) and an autonomous underwater vehicle (AUV) connected by a fixed-length tether. In adjacent reef-monitoring literature, the same name is also used as a design label for broader coral-analysis workflows that combine georeferenced mapping, semantic segmentation, multimodal reasoning, morphology analysis, and restoration handling, so the term now spans both a concrete marine robotics framework and a wider conceptual ecosystem for coral reef observation and intervention (Buchholz et al., 4 Aug 2025).

1. Terminological scope and research context

The explicit CoralGuide formulation is a marine robotics framework centered on safe navigation for tethered ASV–AUV systems in structured marine environments. Its stated purpose is path planning and trajectory optimization, with catenary curve modelling for tether management and Bézier interpolation for smoother trajectories. Several related works, however, are framed as components “for CoralGuide” or for a “CoralGuide-like system,” including georeferenced video mosaicing, coral visual question answering, coral-condition classification, reef metric extraction, and morphology analysis. This suggests that the name has acquired a second, broader meaning as shorthand for an integrated coral-monitoring stack, even when the underlying paper does not itself present a system named CoralGuide (Maypa et al., 1 May 2025, Han et al., 14 Jul 2025).

A common source of ambiguity is therefore categorical rather than technical. In one usage, CoralGuide is a tether-aware planner with explicit robot states, tether constraints, and synchronized trajectories. In the other, it is a conceptual umbrella under which reef informatics modules—mapping, segmentation, VQA, ecological reporting, and restoration tooling—are assembled into an operational workflow. The literature supports both readings, but only the tethered multi-robot framework is a standalone system formally introduced under the name CoralGuide.

2. Tethered ASV–AUV motion planning framework

CoralGuide models an ASV constrained to the water surface and an AUV moving in 3D, with positions

$$P_{\text{ASV}}(t)=\big(x_{\text{ASV}}(t),\, y_{\text{ASV}}(t),\, Z_{\text{ASV}}\big)$$

and

$$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$

subject to the fixed-length tether constraint

$$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$

The planning objective is to generate trajectories for the ASV, AUV, and implicitly the tether such that all remain obstacle-free while respecting tether feasibility over the full mission interval (Buchholz et al., 4 Aug 2025).

Its algorithmic pipeline is hierarchical. First, A* generates an initial path for either the ASV or the AUV. Second, the path for the other vehicle is computed while considering tether geometry and nonholonomic constraints. Third, Bézier interpolation smooths the combined waypoint sequence:

$$B(t)=\sum_{i=0}^{n} \binom{n}{i}(1-t)^{n-i} t^i P_i,\qquad t\in[0,1].$$

Fourth, trajectory optimization and synchronization enforce speed, acceleration, and tether-length constraints. Finally, cubic-spline velocity smoothing prepares the result for PID-based execution. The framework explicitly includes inputs such as ObsMap, TetherLen, MaxSpd, MaxAcc, nonholonomic constraints, and the choice of which robot plans first (Buchholz et al., 4 Aug 2025).

Tether management is encoded through a 3D catenary parameterization between the vehicle endpoints:

$x(t)=x_1+(x_2-x_1)t,\qquad y(t)=y_1+(y_2-y_1)t,$

$$z(t)=a\cosh\!\left(\frac{d(t-0.5)}{a}\right)+z_0-a\cosh\!\left(\frac{d}{2a}\right),$$

with

$d=\sqrt{(x_2-x_1)^2+(y_2-y_1)^2}.$

The tether is discretized into lumped points for obstacle-feasibility checks. Temporal coordination is handled by synchronized waypoint times,

$T_{\text{sync}}=\max_{1\le i\le n} T_i,$

with segment speeds adjusted as

$$V_i=\frac{d_i}{T_{\text{sync},i}-T_{\text{sync},i-1}}.$$

The reported optimization objective also includes terms based on trajectory velocity and acceleration smoothness, while enforcing obstacle-free motion and tether-length feasibility (Buchholz et al., 4 Aug 2025).

Validation included four simulation scenarios—obstacle-free, blowout preventer navigation, wind turbine proximity, and a combined scenario—with ten experiments per scenario. The framework achieved zero collisions in the tested challenging scenarios, while a baseline or ablated A* variant incurred many collisions. Field experiments were conducted in a $12 \times 10 \times 3$ m tank using BlueROV2 platforms, one constrained to act as an ASV, and the reported maximum trajectory deviations were $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$0 mm, $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$1 mm, and $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$2 mm for the ASV along $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$3, $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$4, and $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$5, and $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$6 mm, $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$7 mm, and $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$8 mm for the AUV. At the same time, the paper is explicit that several details are not provided: the exact A* heuristic, admissibility or consistency proofs, a full disturbance model, and a full dynamic tether model remain unspecified (Buchholz et al., 4 Aug 2025).

3. Spatial reef mapping, mosaicing, and georeferencing

CoralGuide-like monitoring workflows are strongly shaped by tools for rapid spatial documentation. CorStitch is a free, open-source software package that automates the creation of georeferenced reef mosaics from underwater video transects obtained through ARRAS-style surveys. Its stitching pipeline uses a Fourier-based image correlation method on a horizontal center strip, operating on the green channel and combining cross-correlation with phase correlation; it exports georeferenced overlays as KMZ files compatible with Google Earth Pro and packages every 100 mosaics into one compressed KMZ file. The method uses a mosaic time of 5 s, a reported strip height of 95 px, and an assumed mosaic width of 3 m for visualization, with timestamp-based matching to echosounder GNSS records (Maypa et al., 1 May 2025).

A higher-fidelity but more computationally intensive path is provided by the photogrammetric workflow that combines underwater imagery, GCP-constrained reconstruction, DSM generation, orthophotos, semantic segmentation, and 2D-to-3D semantic transfer. In the Moorea case study, this produced 1 mm resolution orthophotos and DSMs, dense point clouds with about 10 million points, meshes with about 10 million faces, and spatial accuracy within 4 mm horizontally and 5 mm total error. The semantic component then labeled orthophoto patches and mapped those labels into 3D space, yielding what the paper describes as 3D fine-grained semantic modeling and rugosity evaluation at millimeter accuracy (Zhong et al., 2022).

The literature also contains a notable evidentiary gap. The abstract of the AUV hotspot-mapping paper on coral reefs describes vision-guided autonomous underwater vehicles, photogrammetry, hotspot maps, and rugosity correlation, but the supplied document content is not the target paper and instead contains generic LaTeX rebuttal guidelines. In the supplied material, all domain-specific AUV, detector, photogrammetry, and ecological analysis details are therefore explicitly unspecified (Yang et al., 2023).

Taken together, these works separate two operational layers. CorStitch supports rapid, GIS-compatible reconnaissance from video transects, while photogrammetric semantic modeling supports slower but much more precise structural change analysis. A plausible implication is that CoralGuide-like systems can interleave both modes: fast georeferenced overview products for routine monitoring and periodic high-resolution 3D surveys for structural and semantic change detection.

4. Semantic scene understanding, classification, and interactive querying

Dense scene understanding for coral reefs now rests on several complementary datasets and models. Coralscapes provides the first general-purpose dense semantic segmentation dataset for coral reefs, with 2075 images, 39 benthic classes, and 174,077 expert-annotated masks under a site-disjoint split. It mirrors Cityscapes in structure, and the strongest benchmarked model, SegFormer–MiT-B5 with stronger augmentations, reached 82.761 pixel accuracy and 57.800 mIoU. The dataset deliberately includes operational clutter—fish, divers, transect line, tools, trash—as well as uncertainty-like labels such as background and dark, which is important for any field-deployable CoralGuide-like segmentation stack (Sauder et al., 25 Mar 2025).

YH-MINER extends this by coupling object detection, prompt-guided segmentation, and a Qwen2-VL-based multimodal large model in an “object detection-semantic segmentation-prior input” pipeline. Its detector achieved $$P_{\text{AUV}}(t)=\big(x_{\text{AUV}}(t),\, y_{\text{AUV}}(t),\, Z_{\text{AUV}}(t)\big),$$9, and the prior-conditioned multimodal classifier reached 88.00% genus-level accuracy and 85.65% Macro F1 while extracting ecological metrics such as coral cover, richness, abundance, Shannon-Wiener diversity, Simpson index, and dominant genera by cover. The design choice is to feed segmentation masks and detection-derived priors into the MLLM rather than asking it to interpret the full cluttered scene directly (Wang et al., 28 May 2025).

CoralVQA pushes the same logic into natural-language interaction. It introduces 12,805 real-world coral images, 277,653 question-answer pairs, 67 coral genera, and 16 question dimensions spanning taxonomy, morphology, ecology, and health. InternVL2.5 was reported as the best overall benchmarked LVLM, but the paper also shows that cross-region generalization is poor: nearly all models drop by more than 30% on unseen marine regions, and tasks such as shape, quantity, and bleaching coverage remain difficult. This is an important constraint on any CoralGuide-like conversational interface: natural-language access does not eliminate domain shift (Han et al., 14 Jul 2025).

A separate line targets multi-label condition analysis rather than genus ID. The Indo-Pacific photogrammetry-based study defines an 8-label formulation covering healthy coral, compromised coral, dead coral, rubble, competition, disease, predation, and physical issues across 20,800 image patches. Its best ensemble—Swin-S + Swin-B + EfficientNet-B7—achieved 63.33% match ratio, 86.39% micro F1, and 81.74% macro F1. The paper is explicit that rare stressors are underdetected and that field deployment should use class-specific thresholds and human review for management-critical labels (Shao et al., 2024).

These systems collectively define a semantic ladder. Coralscapes supplies scene-level dense labels; YH-MINER turns detection and masks into ecological metrics; CoralVQA exposes image understanding through VQA; and multi-label condition models target ecological state and stressors. This suggests that a broader CoralGuide can be interpreted as an orchestration layer over segmentation, taxonomic classification, and question answering rather than a single end-to-end model.

5. Morphology, geometry, and biomineral structure

Several neighboring lines of work address structural and morphological analysis at scales ranging from colony architecture to nanocrystallography. CoDA, the Coral Dendroid structure Analyzer, is a visual analytics framework for CT-derived analysis of dendroid cold-water corals. Starting from calyx instance segmentation, it propagates labels to corallites, constructs a region adjacency graph, fits a parabola to each calyx, orients edges into a directed skeleton tree, and enables proofreading, generation assignment, branching analysis, and feature exploration through CoDA.Graph. The resulting representation is explicitly ontogenetic: vertices correspond to corallites, and directed edges encode mother–daughter budding relationships (Schmitt et al., 2024).

At a far smaller scale, x-ray linear dichroic ptychography was demonstrated on Seriatopora aculeata coral skeleton particles as a bulk-sensitive nanoscale crystallographic imaging method. The study mapped aragonite c-axis orientations at about 35 nm spatial resolution, found both narrow $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$0 and wide $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$1 c-axis angular spreads, and identified co-oriented but disconnected ~100 nm subdomains interpreted as evidence of jagged crystal boundaries consistent with amorphous nanoparticle attachment. This is not a CoralGuide module in the operational robotics sense, but it defines the nanoscale material context within which colony-scale geometry and mechanical function emerge (Lo et al., 2020).

A more direct bridge to field geometry is provided by the mesh-free multi-view estimator for coral volume and surface area. That framework uses a pre-trained VGGT point-map module, merges multi-view points into a confidence-enriched cloud, and applies dual DGCNN heads to predict volume and surface area together with uncertainties. On a 15-sample comparison against a mesh-generation baseline, it reported 10.27% volume MAPE and 7.56% surface MAPE, substantially better than a Trellis baseline that yielded 60.81% and 55.41% respectively. Because the model predicts geometry statistics directly from sparse multi-view RGB rather than requiring watertight meshing at inference time, a plausible implication is that CoralGuide-like systems can estimate growth-related quantities under conditions where full 3D reconstruction is brittle or too expensive (Farchione et al., 14 Sep 2025).

These works show that “guidance” in coral research can refer not only to navigation or semantic interpretation, but also to structural inference. In that broader sense, CoralGuide-like analysis spans colony topology, surface geometry, and internal skeletal organization across several orders of magnitude.

6. Restoration, deployment logic, and future directions

Coral-related guidance is not limited to observation; it also includes intervention. ReefFlex addresses safe robotic manipulation of fragile, geometrically heterogeneous coral in cluttered aquaculture environments through generative soft finger design. The paper frames the need against severe reef decline, noting that 13.5% of the world’s hard coral was lost between 2009 and 2018, and that 2024 aerial surveys of Australia’s Great Barrier Reef found 73% of reefs showed bleaching and 39% showed extreme bleaching. Its optimization uses reduced grasping primitives encoded as load cases in density-based topology optimization, and in tight-spacing in-situ tests one optimized finger design, $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$2, grasped 8/10 coral plugs compared with 1/10 for the Fin Ray benchmark (Pinskier et al., 9 Feb 2026).

For mobile monitoring, the CORAL framework—developed for oyster reef monitoring rather than coral reefs—offers a complementary architectural lesson. CORAL decouples high-level semantic reasoning from low-level reactive control: a VLM selects waypoints, a dynamics-based local planner executes them, and a geometric verification module rejects implausible proposals. Compared with the previous state of the art, it improved coverage by 14.28 percentage points, or 17.85% relatively, reduced collisions by 100%, and required 57% fewer VLM calls. A plausible implication is that CoralGuide-like reef robots can benefit from the same semantic/local-planner decomposition rather than relying on end-to-end VLM action generation (Wu et al., 16 Mar 2026).

Strategic deployment is also shaped by ecological prediction. In the East African reef time-series study, among-site variability contributed more to long-term variability in reef composition than temporal variability, with the generalized-variance ratio estimated at $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$3 and site-specific long-term low-coral-cover probabilities ranging from $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$4 to $$l \ge \left\|P_{\text{AUV}}(t)-P_{\text{ASV}}(t)\right\|,\qquad \forall t.$$5. The paper concludes that site identity is highly informative and that potential climate refugia can be identified probabilistically at site scale. This suggests that CoralGuide-like systems are not merely technical platforms but can be targeted toward reefs where long-term coral persistence, restoration need, or monitoring value is highest (Allen et al., 2016).

A final clarification follows from the literature as a whole. CoralGuide is already a concrete tether-aware navigation framework, but it is not yet a single unified platform spanning navigation, mapping, semantic interpretation, morphology, and restoration. The published corpus instead presents interoperable modules: tethered motion planning, georeferenced mosaicing, dense scene segmentation, prior-conditioned metric extraction, VQA, 3D semantic modeling, morphology analytics, and safe coral handling. A plausible future CoralGuide would integrate those modules into one operational stack, but the current literature documents them mainly as linked components rather than as one monolithic system.