Collecting marine data in-person is often infeasible at depth because it's difficult to provide life-support infrastructure in extreme pressure environments. Remotely operated underwater vehicles (ROUVs) offer an alternative, but require continuous communication with a surface operator. Such communication is limited by the rapid attenuation of signals in water and, if instead tethered, by the hydrodynamic drag imposed by any cable. These constraints motivate the use of autonomous underwater vehicles (AUVs) instead, which use onboard computers and decision-making algorithms.
AUVs rely on battery power. This means when designing these algorithms we need to pay careful attention to energy efficiency to avoid frequent retrieval, recharging, and redeployment. Maximizing data-per-unit-energy could involve:
- Using low-energy sensors (e.g., hydrophones instead of cameras).
- Moving the AUV to a better location when replacing the batteries.
- Having the AUV move to a better location itself.
- Use multiple AUVs and have them communicate (minimally).
This repository contains a collection of exploratory notebooks about these options. Together, they amount to a brief study of energy-efficient marine inference.
I focus on a single scenario where you want to conduct a population survey of deep-sea animals, estimating the distribution of species in a target region.
An overview of each notebook is given below.
Today, species identification in the deep ocean is typically performed using video cameras paired with artificial lighting. The system is dropped for a few days and left to record, before being retrieved and analysed. Not only is this approach energy-intensive, but the artificial lighting disturbs some species and alters natural behaviour, biasing observations.
This notebook explores whether passive acoustics can provide a viable alternative. Hydrophones use far less energy than a camera and LED system. The core idea is to use the sound generated by animal movement to infer the species. I create simple synthetic creatures and use their simulated acoustic responses to train a basic classifier as a proof-of-concept. Deployment would involve an initial calibration phase, during which recordings would need to be paired with visual labels, followed by an extended autonomous operation.
Collection of training data would mean navigating to places with high animal density, which is covered in notebooks B, D, and E.
Simulated creatures and their movements within a hydrophone box.
Imagine you are trying to put your AUV at the point of highest animal density, given a set of observations. One option is to use a Gaussian Process (GP) for optimization. The main benefit is the method is very sample-efficient, which means fewer trips. This notebook provides an overview of the method.
Evolution of the GP posterior and during optimization.
Here, I do a slight pivot to study event detection in a network of stationary receivers operating in a noisy environment. For some species like whales observing them close-up is unlikely. Multiple agents recording at different position might be needed to do reliable detection.
I compare:
- A standard centralized aggregation approach, and
- A communication-efficient variant where nodes perform local inference and only transmit when a confidence threshold is exceeded.
The latter requires handling accidental or spurious triggers.
Example of a weak signal embedded in background noise.
Sensor grid layout and corresponding received signals.
Spatial score grid after distributed inference.
This notebook extends the GP-based optimization demonstrated in notebook B by explicitly incorporating energy constraints. While GP optimization is sample-efficient, we also might want to consider the movement cost between locations. While the cost of moving the boat and crew around is mainly determined by the number of samples you take (unless you're travelling many miles), the cost of moving a AUV around underwater is instead dictated by the distance.
I briefly review the existing approach of using mutual information for planning a trajectory and then demonstrate that a simple down-weighting of expected improvement by travel distance can improve the performance of an energy-constrained optimizer.
GP optimization when travel cost is explicitly penalized.
This notebook doesn't address energy efficiency, but does look at how to navigate toward regions with high event rates when all you see are events over time. We also might want to optimise a density of events such as animal observations in order to collect lots of data for something like the classifier described in notebook A.
The problem can be framed as inference over an inhomogeneous Poisson process, or a kind of continuous multi-armed bandit problem. I test an approach where observed events are converted into an events-per-unit-time signal along the robot's path, which is then used to reconstruct the full spatial intensity function.
True underlying event intensity function.
Observed events over time along the robot’s trajectory.
Reconstructed spatial intensity function from sparse event data.