Active localization on the ocean floor with multibeam sonar

We present a method for infrastructure-free localization of underwater vehicles with multibeam sonar. After constructing a large scale (4 km), high resolution (1 m) bathymetric map of a region of the ocean floor, the vehicle can use the map to correct its gradual dead-reckoning error, or to re-localize itself after returning from the surface. This ability to re-localize is particularly important for deep-operating vehicles, which accumulate large amounts of error during the descent through the water column. We use a 3D evidence grid, stored in a efficient octree data structure, to fuse the multibeam range measurements and build maps that do not rely on particular features and are robust to noisy measurements. We use a particle filter to perform localization relative to this map. Both map and filter are general, robust techniques, and both run in real-time. Localization convergence and accuracy are improved, particularly over sparsely varying terrain, by deliberately selecting actions that are predicted to reduce the vehicle's position uncertainty. Our approach to this active localization is to select actions that are expected to generate sonar data that maximally discriminates between the current position hypotheses. Maximal discrimination is a very fast proxy for standard particle filter entropy-based active localization. These methods are demonstrated using a dataset provided by the Monterey Bay Aquarium Research Institute from their mapping AUV, collected near the Axial Seamount in the Juan de Fuca Ridge. Though it depends on the situation, the vehicle's position estimate typically converges to within 2 m of the true position in less than 100 s of traverse, or 150 m at 1.5 m/s. We explore the limitations of our approach, particularly with a smaller number of range sensors: although performance is degraded, satisfactory results are achieved with just four sonars.