SOLAQUA: SINTEF Ocean Large Aquaculture Robotics Dataset
Sveinung Johan Ohrem 1,† , Bent Haugaløkken 1 , Eleni Kelasidi 1,2 · 2025
Read SOLAQUA: SINTEF Ocean Large Aquaculture Robotics Dataset by Sveinung Johan Ohrem 1,† , Bent Haugaløkken 1 , Eleni Kelasidi 1,2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-01790.txt.
Introduction
Sveinung Johan Ohrem, Bent Haugaløkken, and Eleni Kelasidi present a data set in twenty twenty-five. The data came from an underwater robot working at a real sea-based fish farm. It was gathered for SINTEF Ocean and named SOLAQUA. The robot moved through a net pen, at times steered by hand and at times on its own.
Sensors on board recorded sonar, cameras in single and stereo pairs, and motion and depth data. Other records cover power use, pressure, and temperature. The nets filmed were undamaged, with some fish and sea growth present. The authors expect both researchers and the fish farming industry to gain from the data. It gives robot teams real farm conditions to train and test against.
Catalog record
| CREATOR | Sveinung Johan Ohrem 1,† , Bent Haugaløkken 1 , Eleni Kelasidi 1,2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-01790
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SOLAQUA: SINTEF Ocean Large Aquaculture Robotics Dataset Sveinung Johan Ohrem 1,† , Bent Haugaløkken 1 , Eleni Kelasidi 1,2 Abstract This paper presents a dataset gathered with an underwater robot in a sea-based aquaculture setting. Data was gathered from an operational fish farm and includes data from sensors such as the Waterlinked A50 DVL, the Nortek Nucleus 1000 DVL, Sonardyne Micro Ranger 2 USBL, Sonoptix Mulitbeam Sonar, mono and stereo cameras, and vehicle sensor data such as power usage, IMU, pressure, temperature, and more. Data acquisition is performed during both manual and autonomous traversal of the net pen structure. The collected vision data is of undamaged nets with some fish and marine growth presence, and it is expected that both the research community and the aquaculture industry will benefit greatly from the utilization of the proposed SOLAQUA dataset. I Introduction Aquaculture is and will be an important contributor to the production of protein and food in the years to come. Aquaculture as an industry is found all over the world, with varying production methods dominating depending on the country [ 1 ] . In Norway, sea-based aquaculture is the current industry standard, as the Norwegian geography with sheltered fjords invites this production type. Atlantic salmon ( salmo salar ) is the most commonly farmed species and the production method of choice is net pens. These net pens, and the surrounding infrastructure, are in constant need of inspection, maintenance and repair (IMR) operations. The aquaculture industry in Norway has adapted underwater remotely operated vehicles (ROVs) for majority of the inspection and intervention tasks in fish farms. Typical ROV operations in aquaculture range from net and mooring inspections and cleaning, to fish monitoring and net pen installations [ 2 ] . A net pen offers a unique domain for underwater robots: The operations are normally performed in - or close to - the splash zone with currents and waves affecting the robot’s states. The environment is dynamic, with large moving structures and several potentially moving obstacles. There are hundreds of thousands of fish in a net pen which will interfere sensor signals and block camera views. All of this combined with the risks and costs of performing sea-based operations makes aquaculture a challenging domain for underwater robotics which is likely to benefit greatly from the development of higher levels of autonomy [ 2 ] . Research efforts have been made over the last years to advance the level of autonomy in the mentioned areas as these operations are mostly manually performed today [ 3 ] . Early results include novel methods for navigation in net pens using a Doppler velocity logger (DVL) and ultra-short baseline (USBL) system [ 4 ] . This work led to the development of a net-following algorithm [ 5 ] which has been demonstrated in several later works, e.g., [ 6 , 7 , 8 ] . Navigation using low-cost sensor systems was investigated in [ 9 ] where various low-cost sonar and DVL technologies are shown to be suitable for navigation purposes in net pens. Camera-based methods are investigated, for example, in [ 10 , 11 ] with the goal of reducing the number of navigation sensors needed on underwater vehicles in aquaculture. (a) ROV with sensor configuration 1. (b) ROV with sensor configuration 2. Fig. 1 : ROV with two different sensor configurations. As methods driven by artificial intelligence (AI) and machine learning (ML) are adapted in the aquaculture robotics domain, the need for datasets to train AI and ML models is of great importance to contribute and advance the research on robust methods for fully autonomous IMR operations, and thus address current and future challenges of the aquaculture industry. Particularly in the field of vision-based and multi-modal navigation, where camera images paired with e.g., position measurements from other sensors such as DVL and USBL can have significant value. To this end, we present the SINTEF Ocean Large Aquaculture Dataset, SOLAQUA, consisting of navigation sensor data, sonar images, mono and stereo camera images, and vehicle data, all captured in a fully operational fish farm. The dataset was captured while performing manual control of a remotely operated vehicle (BlueROV2) as well as during autonomous traversal of the net structure using the previously developed net following algorithm. The data may be utilized for the development of navigation methods, developing novel computer vision techniques for navigation and inspection, simulation purposes, to verify novel control methods, and much more. Examples of use of the dataset for development of robust localization and mapping methods for UUVs operating in industrial scale fish farms can be found in e.g., [ 12 ] . The dataset consists of numerous subsets of data from an ROV operating between 1 and 2 minutes at different depths, with various speeds, net-relative distances, and facing different environmental conditions. The ROV faces the net structure during the runs, hence images are of undamaged nets with some marine growth. Some fish are present, but the purpose of this dataset is not to capture images of fish. To our knowledge, no similar dataset exist. This paper briefly describes the platform used for gathering the data (a BlueROV2) and the data acquisition procedure. In addition, information are presented to give detailed overview of the sensors, their logging frequency and position on the ROV. This overview together with the concrete examples of data found in the dataset can be utilized from the research community to develop and propose novel methods. The dataset can be found at the following website: https://data.sintef.no/feature/fe-a8f86232-5107-495e-a3dd-a86460eebef6 and is under CC BY-SA license. II Multi-modal underwater platform for data collection and autonomous operations In the presented dataset the vehicle was either moved manually or autonomously using net following. Net following refers to an automatic control procedure where the vehicle moves relative to the net at a desired distance, heading angle, and velocity relative to the net structure [ 5 ] . All datasets are logged with time-synchronization using ROS (Robot Operating System), i.e., the time-stamp of each logged data point is generated by the same clock. Due to varying data capturing frequencies between the sensors and generated signals, minor variations in time-stamps, in the millisecond range, are present. No known delays occurred while capturing the data. The data was logged as .bag-files. As shown in Table I and Figure 1 two configurations of the BlueROV2 was used to gather the datasets. Each configuration results in a vehicle with different dimensions and weights, and with a variety of integrated relevant sensors. An overview of the ROV system architecture with detailed diagrams illustrating the connections between sensors, actuators, control system and communication interface can be found in [ 9 ] . TABLE I : BlueROV2 Technical Specifications Feature Description Size (L x W x H) (Config 1) 457 x 436 x 397 [mm] Size (L x W x H) (Config 2) 457 x 436 x 541 [mm] Weight in air/water (Config 1) 15/0 [kg] Weight in air/water (Config 2) 18/0 [kg] Number of thrusters 8 Communication Ethernet, UDP Degrees of Freedom 6 Power Battery, 14.8 [V] II-A Sensors A list of sensors used to gather the datasets is shown in Table IV . Note that some custom ROS messages have been used.