Feature Space Exploration For Planning Initial Benthic AUV Surveys

TL;DR

The paper addresses the challenge of planning informative initial benthic AUV surveys over vast seafloor areas by exploiting bathymetric priors to sample a feature space representing seabed terrain. It formalizes an offline framework that links feature-space exploration to physical-path planning, introducing an information-reward based on the Mahalanobis distance and evaluating planners (RRT, MCTS, Template placement, and Cluster-TSP) against metrics that quantify feature-space coverage and habitat visitation. The study demonstrates that RRT and MCTS-based planners substantially improve feature-space exploration over random sampling across multiple feature representations (geometric and learned). These methods increase the utility of initial dives and provide rich training data to relate acoustic bathymetry to visually-derived seafloor classes, with field trials highlighting practical considerations and the potential of shorter informative surveys under real-world constraints.

Abstract

Special-purpose Autonomous Underwater Vehicles (AUVs) are utilised for benthic (seafloor) surveys, where the vehicle collects optical imagery of the seafloor. Due to the small-sensor footprint of the cameras and the vast areas to be surveyed, these AUVs can not feasibly collect full coverage imagery of areas larger than a few tens of thousands of square meters. Therefore it is necessary for AUV paths to sample the surveys areas sparsely, yet effectively. Broad-scale acoustic bathymetric data is readily available over large areas, and is often a useful prior of seafloor cover. As such, prior bathymetry can be used to guide AUV data collection. This research proposes methods for planning initial AUV surveys that efficiently explore a feature space representation of the bathymetry, in order to sample from a diverse set of bathymetric terrain. This will enable the AUV to visit areas that likely contain unique habitats and are representative of the entire survey site. We propose several information gathering planners that utilise a feature space exploration reward, to plan freeform paths or to optimise the placement of a survey template. The suitability of these methods to plan AUV surveys is evaluated based on the coverage of the feature space and also the ability to visit all classes of benthic habitat on the initial dive. Informative planners based on Rapidly-expanding Random Trees (RRT) and Monte-Carlo Tree Search (MCTS) were found to be the most effective. This is a valuable tool for AUV surveys as it increases the utility of initial dives. It also delivers a comprehensive training set to learn a relationship between acoustic bathymetry and visually-derived seafloor classifications.

Figures (24)

• Figure 1: An overview of linked feature space and physical space exploration, where the objective is to comprehensively explore the feature space given a budget in physical space. Features extracted from along the survey path are projected into the feature space. In this demonstration, the blue squares indicate features already on the path, with the planner making a decision about where to move physically (indicated by crosses). The planner chooses the green cross over the red cross as it occupies a more unexplored area of the feature space.
• Figure 2: The VAE used for feature extraction. Square ($21\times21$ pixels) patches of bathymetry are extracted from random locations of the bathymetry. These patches are compressed into a representative feature space by the autoencoder, before being reconstructed by the decoder aiming to recreate the sample patch. Using helps generate a continuous feature space, which is useful for exploring.
• Figure 3: An overview of the process for planning an informative initial survey. The inputs to this process are the survey area and corresponding remotely-sensed data. Using the remotely-sensed data, the planner designs a survey plan that uniformly samples from the feature space. The output of this process is a set of waypoints that the AUV is tasked to follow. Finally, the AUV is deployed and captures benthic seafloor imagery.
• Figure 4: Visualising the process for evaluating a path. On the left it shows the physical (spatial) space and the robot's path. The green crosses are points sampled on the robots path, while the red crosses are points randomly sampled from the survey area. The right plot shows the robot's path in feature space. The red crosses are clustered in feature space to form the $k$ centres. For each of the $k$ centres, the distance to the closest feature from the path is found. The total evaluation score is the sum of all these distances.
• Figure 5: InfoRRT Process. First, the planner selects a target point from the larger survey area. This target point is selected either at random or is the highest reward point from many candidate target points. Next, nodes on the search tree that are closest to the target point are evaluated for expansion. This node is selected such that it will be the highest value, when combined with the target point. A new node is created by steering the node in the direction of the target point, and finally this node is added to the search tree.