On January 10th 2008, the Autonomous System group working with John Ryan, undertook a full day cruise to demonstrate in-situ deliberation coupled with online classification for Intermediate Nepheloid Layer (INL) mapping. T-REX, the onboard deliberative control system, on MBARI's CTD AUV platform was coupled with our supervised cluster learning algorithm. In the same spirit as the Elkhorn Slough plume mapping, our goal was to change dynamically the spatial resolution of the volume survey based on the INL signal strength end fire adaptively the Gulper to retrieve gulper samples when the signal was strong enough. To detect the plume we used data from 6 previous missions to learn features associated with an INL. The resulting probabilities were between 0 and 0.99. We decided to fire the gulper when the expected probability was greater than 0.4 and to change the survey resolution from 250m for a very strong signal to 750m with no signal detected.
The data below summarizes the two most representative runs on this day.
Science mission :
We had some difficulties during the mission. First the 4 first transects were done in high resolution because the classification was based on buggy data resulting on a returned probabilty which was always around 0.8. As a consequence our first 6 samples were taken almost randomly and the vehicle made this transect with a high resolution (250m). After correction during the mission, with the vehicle on the surface, we relaunched the mission which behaved as expected (spatial separation around 400m with an obvious INL signature all alonfg the tarnsect).
Even though, the adaptation to the signal was good we had two other issues to deal with :
- The classification was not as accurate as we expected based on our on-shore tests and the probability threshold of 0.4 was too low to be really associated with a plume. As a result T-REX fired the gulper too soon and missed the most relevant parts of the water column. We should have ever selected a higher threshold -- which appeared to us as risky before the run -- or have a better policy which would look more to the local maxima to be sure to fire the Gulper with a stronger signal
- We also saw a situation indicating a temporal inconsistency in the T-REX model. The safety control added to allow T-REX to deliberate without allowing the vehicle to be uncontrolled worked well.
Because of these 2 problems the mission was split into 3 smaller missions allowing us to restart quickly without complex modifications at sea. The figure below shows the resulting data used for INL detection and where the samples were taken. The 3 mission are merged to give the reader a global view of this trial. On this figure the p(INL) graph represents the probabilty to be in the INL based on classification output. This value was corrected for first trial.Processed science data
INL mapping : classification result and gulper firing.
The water samples taken are being analyzed to date by scientists. From our side we will need to correct the model inconsisitency and improve our INL cluster model and to have a better policy when to fire the gulper.