2024 Autonomous Systems Bootcamp

Ocean gliders are autonomous underwater vehicles that have become invaluable tools for scientists, providing them with detailed information about oceanic conditions over large spatial scales and extended periods, which is crucial for understanding climate dynamics, marine ecosystems, and various oceanographic processes. They may be equipped with a wide variety of sensors to monitor temperature, salinity, currents, and other ocean conditions. This information creates a more complete picture of what is happening in the ocean and trends scientists might not otherwise be able to detect from satellites or large research ships (source: NOAA Ocean Service). 

The glider’s ability to change buoyancy allows it to dive and ascend through the water column in a controlled manner, while the hydrofoils help convert some of this vertical motion into horizontal movement. This method of propulsion is highly energy-efficient, making underwater gliders capable of extended missions lasting weeks to months. However, this type of motion makes such vehicles susceptible to environmental conditions such as ocean currents and subsequently their trajectory control challenging. What would be helpful to the scientists piloting gliders is a decision support tool that outlines the possible glider trajectories, given encountered ocean conditions (acquired on a broad scale via e.g. satellites). 

This project aims to design an estimation model predicting the glider’s position in a 3-day window. Given the available prediction for environmental conditions for that window and knowing the vehicle dynamics and its current position, we can estimate its potential trajectory. Such a solution will allow the pilots to optimize glider trajectories while increasing the potential range of locations sampled and allowing them to take corrective actions to increase the safety and positional accuracy of the glider’s missions.

There are multiple reasons why we care about oysters. Apart from the fact that, according to NOAA, shellfish production represents a large and growing segment of the United States and global seafood industry, oysters are filter-feeders, removing algae, organic matter, and excess nutrients from the water column as they grow and improving water quality. They form beds that stabilize coastal sediments and help minimize impacts from storm surges, and their production through oyster farms has a benign ecological footprint with little disturbance of sediments or aquatic vegetation during grow-out. (source: NOAA fisheries) 

One of the challenges of on-bottom oyster farming is the need for consistent and efficient monitoring of the farm. Traditional sampling methods are labor-intensive and often destructive. To this end, we are interested in exploring how we can effectively monitor the condition of the oysters in on-bottom aquaculture farms using robotic platforms and machine learning.

The goal of this project is to design a DNN (Deep Neural Network) for oyster detection and identification through optical sensor data. The oysters should be detected and identified as open/closed and dead/alive. The first milestone of the project is to achieve the highest accuracy while running offline; however, the final solution should also include a “lighter version” to run in real-time on the Aqua 2 robotic platform. Diverse datasets will be provided, including footage collected using different platforms (BlueROV, Aqua 2, GoPro). Such solution will (i) provide estimates for both biodiversity and abundance at multiple depths, and (ii) increase efficiency and decrease difficulty of continuous monitoring that may provide a more complete picture than “point in time” techniques.

Object detection in aquatic environments has been a major topic of attention in recent years, mainly focusing on detecting different elements in the water using optical data from various types of cameras. However, in areas with low visibility in the water, well-known object detection techniques through optical sensors are not an option. An alternative to optical sensors that can work more efficiently in turbid water is using acoustic sensors, like sonar. Such sensors have been broadly used in the marine environment for several years, yet not much work has been done on using acoustic data to train machine-learning models for marine applications. The goal of this project is to design a neural network using acoustic data that can accurately detect static elements in the water, such as derelict crab pots, and moving elements, such as animals. There are two main target applications in which such a solution can be essential.

In the first application, we are interested in detecting derelict crab pots located in the Delaware Inland Bays. The shallow, protected habitats of Delaware’s Inland Bays make for one of the most popular areas in the state for residents and tourists to try their hands at catching blue crabs. In boats or on the shore, recreational crabbers use all kinds of gear, from hand lines to trot lines to small traps with collapsible sides and the Chesapeake-style crab pot. Unfortunately, thousands of derelict crab pots have been left behind or lost and are littered beneath the surface of the Inland Bays (source: NOAA). In addition to littering the seafloor, these derelict crab pots can cause damage to boat propellers and have the potential to ‘ghost fish,’ luring sea creatures into their midst and trapping them. Common species encountered in the pots are diamondback terrapins, blue crabs, and oyster toadfish (source: Delaware Sea Grant). Thus, detecting those crab pots in a more accurate and time-efficient way is very crucial.

For the second application, we are interested in employing a machine-learning model to measure the efficiency of different fishing techniques. More specifically, we are interested in estimating the number of animals present in the area while we place fishing gear in the water. Knowing this number will allow us to compute a more accurate catch rate, which can also provide essential insights about the total animal population of the area. A commercially important fishing technique that we are interested in is the so-called longline fishing. In this method, we use a long main line with baited hooks that can be floated at variable distances below the surface and lies more or less horizontally in a series of sagging connections between each buoy, with numerous short hooked lines attached and hanging below (source: NOAA fisheries). 

One of the most common ways to gather information about the movement and behavior of marine organisms is by using animal-borne sensors or tags. Over time, continuous, long-term observations illustrate not only animal movements but also help us see the signs and understand the effects of changes to the ecosystems they inhabit. For this project, we are interested in the so-called pop-up archival transmitting (PAT) tags, which are externally placed tags pre-set to detach and rise to the surface after the data collection (source: Census of Marine Life). After catching the animal, the tag is inserted on its fin and remains on the animal for a predetermined time. After this time period passes, the tag detaches itself and sends a signal to the scientist team, initiating their search to reacquire the equipment. The tag will continue sending location signals irregularly till it’s recovered. 

The two primary tools utilized to search the detached tag are a hydrophone antenna providing direction and a bandwidth antenna providing proximity. The reference zero angle of the hydrophone is aligned with the bow of the search vessel, and every time there is a new signal from the tag, the vessel is steered in that direction. Since such a process can be time-inefficient, the goal of this project is to develop a solution to decrease the time of recovery by incorporating a robotic platform in the search process. Adding a second hydrophone antenna on an autonomous surface vehicle (ASV) can potentially accelerate the search by using the two angles (one from the support vessel and one from the ASV) to compute the exact location of the tag via triangulation. 

Multibeam sonar is an active sonar technology utilized for mapping the seafloor and detecting objects within the water column or along the seafloor. It operates on the principle of reflecting sound off the seafloor in a fan-shaped pattern and is an efficient way to systematically map large regions.  Multibeam sonar allows scientists to generate a point cloud of the seafloor (bathymetry) as well as capture textural information (backscatter) using the two way travel time.  This allows them to create detailed maps of the underwater terrain for studying underwater geological processes, mapping hydrothermal vents, identifying potential seafloor habitats, and locating submerged cultural heritage sites. Analyzing backscatter data can help scientists classify seafloor habitats and identify geological features. (source: NOAA Ocean Explorer)

The coverage of the multibeam sonar swath varies directly with the altitude of the sensor (equal to depth on a surface vessel). On a seafloor with changing depths, the swath coverage can change significantly over the course of a survey. What is the optimal path of a vessel or ASV that is surveying such an area given the constraints of the navigation and propulsion equipment onboard?

The aim of this project is to optimize the time of mapping survey missions and subsequently, the cost associated with them by adjusting the shape and spacing of the survey lines based on depth and without compromising the quality of the data acquired. This problem can be split into two subproblems (i) Design the lines of the survey before the mission based on prior depth information about the mapping site and (ii) design the lines of the survey while being on mission, assuming that there is no information about the depth of the site and online path planning based on real-time depth-information is necessary.