Cooperative Terrain Relative Navigation PDF Download

Terrain-Relative Navigation (TRN) is an emerging technique for localization in natural environments. TRN augments a dead-reckoned solution with position fixes based on correlations with pre-stored maps. TRN is a particularly valuable tool for enabling missions for robots in regions without GPS, a category that includes the underwater environment as well as missions on other bodies in the solar system. The algorithms underlying TRN, however, have known issues with overconfidence in uninformative (e.g. flat) terrain. Overconfident estimates, also known as false peaks, are a significant problem as they can result in dangerous trajectories and mission failure. Making TRN robust to uninformative terrain is the focus of the work presented in this thesis. The interplay between map error, terrain correlation, and TRN filter overconfidence is the first focus of this thesis. TRN correlation techniques are shown to include, either implicitly or explicitly, a probabilistic model of terrain correlation, and that the most common method of TRN weighting implicitly models the terrain as uncorrelated. The degree of auto-correlation present in the terrain is related to the amount of variation in the terrain: greater variation in terrain height corresponds to lower correlation in the terrain and vice-versa. The uncorrelated terrain assumption is then demonstrated to be a source of false peaks. In informative terrain, where the variation in the terrain is large with respect to error in the map, the standard calculation produces reasonable results: peaks at the correct location. In uninformative terrain, when the variation in the terrain is small with respect to map error, standard correlation breaks down and is shown to produce overconfidence in the filter. Techniques are then developed for mitigating false peaks in uninformative terrain. The first technique developed in this thesis focuses on explicitly accounting for terrain correlation by using correlated Gaussian terrain models; while these methods have success in simulation, the computational cost of explicitly modeling terrain correlation makes them impractical for field applications. An alternate approach, exponentially down-weighting the standard weighting to account for the impact of the uncorrelated terrain assumption, is then proposed as a computationally tractable means of accounting for unmodeled terrain correlation. The exponential down-weighting technique is termed the adaptive TRN filter. It follows on work from the statistics community designed to improve the robustness of probabilities computed using incorrect models, and achieves this robustness by matching a bound on the likelihood of false peaks. The ``robust adjusted likelihood'' approach is adapted to the TRN likelihood function and used to develop the relation between terrain correlation and the necessary degree of down-weighting. The adaptive technique is further developed for field work using real-time TRN filters. The adaptive TRN filter is validated using two platforms: an Autonomous Underwater Vehicle (AUV) and an ATRV-Jr ground rover. The AUV TRN filter is developed for an AUV correlating with range measurements of the terrain. The ground rover filter is developed for operations on the Moon or Mars, where direct measurements of altitude are unavailable, and the TRN filter must therefore correlate on gradient. As most maps are elevation based and must be differentiated to produce a gradient map, the map noise is increased and makes accounting for map error critical in this case. The effectiveness of the adaptive TRN filter is demonstrated using field data from MBARI AUV runs over flat terrain in Monterey Bay, and on ATRV-Jr field data taken at the Stanford campus. Both cases demonstrate meter-level performance when operating in informative terrain, and effective mitigation of false convergence over uninformative terrain when compared to filter performance using the unadjusted weighting.