Monterey Bay Aquarium Research Institute
Autonomy
7 hours of underwater adaptation
Long duration adaptative survey


November 10th 2008 :

On this cruise, our objective was to demonstrate that T-REX was robust enough for a long (> 4 hour) mission and accurate in its estimation. While the mission was similar to  Jan. 10th there were several improvements in our TREX domain model as well as the classifier used to identify the INL. Once again we deployed the AUV in waters of northern Monterey Bay.

During the uninterrupted 6 hours and 40 minutes run  T-REX was able to identify the INL and sample it accurately with the Gulper instrument. As shown in the figure on the right, the vehicle started its survey within an INL (yellow) and ended up in a region with a weak signal. It was able to change its survey resolution (transect spacing) accordingly starting with high resolution transects and ending with a low resolution survey.

Sampling was undertaken  in the neighborhood with a strong estimate (see figure below on the left). However the classifier still came up short by sampling at the boundaries of the INL instead of where the INL signal was the strongest.




We improved this estimation using a new incremental learning technique that tries to insert new clusters in the previous model by reduceing  relative entropy using newly collected data. This resulted in a vastly more precise model (see figure below with the updated model) showing sampling closer to the centroid of the INL.

After a quick implementation of this estimation model, we were  ready for the next sea trial just 3 days later.

November 13th 2008:

There were two issues in this cruise:

  1. The vehicles functional layer encountered a previously unseen bug not actuating the propulsion  after 3 hours of correct execution. T-REX itself performed flawlessly. This resulted in a two-part 7 hour mission.
  2. In trying to reduce the new estimation model size we had inadvertantly produced a model which could generate false positives (samples taken while not within an INL). This problem appeared during the second run with 2 samples made close to the surface while others samples were taken inside INL hotspots.
T-REX behaved as expected and was able to smothly recover from the vehicle anomalies. Processed data for the two runs is provided below.

The false negatives were due to an attempt to reduce model size using an empirical approach with this being our first attempt in model improvmentl incrementally. The initial result was a larger model; consequently in trying to remove less significant parts of the model empirically a critical cluster of the model was removed. 

Last updated: Feb. 06, 2009