cs.ROSep 30, 2026

Coral Grow-out Robotic Assessment System (CGRAS): Scaling Coral Recruit Monitoring Through Robotics and Computer Vision

Authors: Dorian Tsai, Scarlett Raine, Emilio Olivastri, Riki Lamont, Andrew Lui, Timothy Morris, Joshua Esplin, Christopher A. Brunner, +7 more

Organizations: Queensland University of Technology (QUT), Brisbane, Australia · Australian Institute of Marine Science (AIMS), Townsville, Australia · National Sea Simulator at AIMS

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Climate change is the largest threat to coral reefs, with increasing global impacts accelerating the need for scalable reef restoration technologies. Large-scale reef restoration depends on the mass production of corals, such as through coral aquaculture. Coral seeding with recruits grown in aquaculture facilities is a feasible restoration approach, but effective production requires consistent, high-frequency monitoring of tens of thousands of macroscopic (0.5-2mm diameter) recruits, making conventional manual assessment prohibitively labor-intensive. To address this monitoring bottleneck, we introduce the Coral Grow-out Robotic Assessment System (CGRAS) which combines robotic imaging and computer vision to automate data acquisition, perform multi-species detection and counting of corals, and evaluate coral health. CGRAS automatically extracts coral growth, survival and spatial distribution metrics, with the aim of providing timely feedback to operators for optimizing production, grow-out and deployment workflow processes. We demonstrate CGRAS in a large aquaculture facility on standardized coral settlement tiles, reducing the time and labor costs by a factor of 9.6 as compared to manual monitoring, whilst achieving 96.4% agreement for Acropora kenti corals relative to expert counts.

Figures & tables

Fig. 1: Coral Grow-out Robotic Assessment System (CGRAS). (a) Image capture system. (b) Magnified example of coral recruit detections, where the red lines denote the 14×1414\times 14 mm tabs that the 280×280280\times 280 mm coral settlement tiles are broken into prior to deployment.
Fig. 2: Coral grow-out starts when competent coral larvae settled on provided concrete tiles (t 0 ). During operation, CGRAS performs continuous monitoring of coral survival (t 1 -t 2 ), ranging from 2 to 12 weeks, enabling real-time optimization of tank conditions. When coral recruits have reached an appropriate growth stage for reseeding onto the reef (t 2 ), the CGRAS heatmaps are used to select the specific tiles and tabs for outplanting.
Fig. 3: The IAS captures two coral imaging tanks in a single operation. Calibration tags establish the spatial relationship between the robot arm base and the imaging tanks, allowing accurate localization of coral tile positions within tanks.
Fig. 4: Examples of coral development for Aken , Amil and Pdae coral recruits. During Weeks 0–1, recently settled corals begin forming skeletal structures that are only faintly visible. Beyond Week 1, skeletal development becomes more pronounced, while the central polyp becomes progressively darker due to the increasing abundance of symbiotic Symbiodiniaceae algae within the coral tissue. Coral mortality is also seen after Week 4 for Aken , and Week 12 for Amil , where soft tissue degradation is followed by overgrowth of algae and other benthic organisms on the exposed skeleton. Pdae are smaller, and largely transparent until Week 4 when they grow in size and visually differ from Aken and Amil .
CategoryMetricValueUnit
Acquisition: IASTiles imaged183tiles
Non-recoverable errors1tile
Imaging time14.6hours
Processing: CCVSTotal tiles182tiles
Successfully processed tiles160tiles
Processing time7.6hours
TABLE I: Operational performance of the CGRAS system, evaluated during 2025 spawning season. See Sec. V-A for details.
Fig. 5: Performance of CCVS object detection models for the coral species Aken . (a) mAP50 as a function of the number of labeled training images, averaged across nine independent training runs with different random seeds and data shuffles (refer to Sec. IV-D for metric defintions). Performance increased rapidly up to approximately 30 labeled images before plateauing, indicating that 30 images provide a practical balance between annotation effort and model performance in data-constrained settings. (b) Precision–Recall curve for the alive coral class of the final model, trained using all 130 labeled images. Although most gains were achieved with relatively few labeled images, the final model used for ML-assisted labeling was trained on the complete dataset to maximize performance.
SpeciesAkenAmilPdae
Aken96.391.164.1
Amil82.396.260.2
Pdae50.255.492.3
TABLE II: Species-specific performance (mAP50; see Sec. IV-D for further details) of the CCVS object detectors. Columns correspond to single-species detectors, while rows correspond to single-species test datasets. Each dataset contains 130 images of live corals from 2024. Best result per column indicated in bold .
Fig. 6: Outputs from the CCVS subsystem. (a) Alive corals (indicated by green squares) detected on a stitched tile image, composed of 24 stitched images. The 20×2020\times 20 grid of tabs is overlaid in red. The black corner tile spacers are detected and used to orient the tile within the stitched image. (b) The corresponding coral heatmap (with absolute coral quantities) with respect to the tab grid, where brighter colors represent a greater number of live corals.
Fig. 7: Distribution of the differences between Aken manual vs CGRAS-based counts per tab. The distribution has a mean of 0.15 miscounted corals per tab, and a standard deviation of 1.59, demonstrating a strong agreement between manual and automated counts.
Fig. 8: Number of live corals on a tile of Aken from 2025, demonstrating the correspondence between the manual counts (blue) and the CGRAS counts (orange).

Explore similar work

CardsList

  1. AI-driven Dispensing of Coral Reseeding Devices for Broad-scale Restoration of the Great Barrier Reef

    Aug 31, 2025Scarlett Raine, Emilio Olivastri, Benjamin Moshirian +1CoralUnderwater Object Detection

    QUT Centre for Robotics and School of Electrical Engineering and Robotics, Queensland University of Technology, Brisbane, Australia · Australian Institute of Marine Science, Townsville, Australia

  2. CoralscapesV2: Panoptic and Fine-Grained Visual Scene Understanding in Coral Reefs

    Sep 14, 2026Jonathan Sauder, Thomas Ruckli, Gabrielė Strodomskytė +17CoralPlankton Classifier

    MIT · EPFL · University of Djibouti +7

  3. CoralLite: μCT Reconstruction of Coral Colonies from Individual Corallites

    May 14, 2026Jess Jones, Leonardo Bertini, Kenneth Johnson +2Coral

    University of Bristol · University of Liverpool · Natural History Museum