||Automated Planning for Hydrothermal Vent Prospecting Using AUVs
||Wed 18 Jan 2012
||Dr Zeyn Siegol
University of Birmingham
||In this talk I will present several decision-theoretic algorithms for planning a search strategy for an AUV. The algorithms are targeted at prospecting for hydrothermal vents, which are superheated outgassings of chemical-rich water that are found at the bottom of the oceans. The ability to find multiple vents efficiently using an AUV is very valuable for oceanographers, and achieving this requires on-board data analysis together with planning under uncertainty.
To find the best possible plan, the AUV should build a probabilistic map of its environment, marking the areas where vents are likely to be found. I use Michael Jakuba's occupancy grid method to create the map, and this map forms the input to the novel planning algorithms. These algorithms have been tested in simulation and found to be twice as effective as using a moving-the-lawn search pattern, and 50% more effective than a biologically-inspired chemotaxis algorithm that does not make use of a map.
I will present the formal description of the problem domain as a partially-observable Markov decision process (POMDP), and two families of planning algorithms. The first family uses online POMDP methods, and works by doing a limited forward search in the action-observation space of the problem. The second family uses map entropy as a heuristic to drive exploration, and is found to be as effective as the more principled but (often) computationally more expensive online POMDP family.
<< Back to list of seminars