Automated planning for hydrothermal vent prospecting using AUVs

This thesis presents two families of novel algorithms for automated planning under uncertainty. It focuses on the domain of searching the ocean floor for hydrothermal vents, using autonomous underwater vehicles (AUVs). This is a hard problem because the AUV’s sensors cannot directly measure the range or bearing to vents, but instead detecting the plume from a vent indicates the source vent lies somewhere up-current, within a relatively large swathe of the search area. An unknown number of vents may be located anywhere in the search area, giving rise to a problem that is naturally formulated as a partially-observable Markov decision process (POMDP), but with a very large state space (of the order of 10123 states). This size of problem is intractable for current POMDP solvers, so instead heuristic solutions were sought. The problem is one of chemical plume tracing, which can be solved using simple reactive algorithms for a single chemical source, but the potential for multiple sources makes a more principled approach desirable for this domain. This thesis presents several novel planning methods, which all rely on an existing occupancy grid mapping algorithm to infer vent location probabilities from observations. The novel algorithms are information lookahead and expected-entropy-change planners, together with an orienteering problem (OP) correction that can be used with either planner. Information lookahead applies online POMDP methods to the problem, and was found to be effective in locating vents even with small lookahead values. The best of the entropy-based algorithms was one that attempts to maximise the expected change in entropy for all cells along a path, where the path is found using an OP solver. This expected-entropy-change algorithm was at least as effective as the information-lookahead approach, and with slightly better computational efficiency.