On November 21st 2007, the Autonomous System group working with John Ryan, undertook a full day cruise to demonstrate in-situ deliberation coupled with online classification for plume mapping. T-REX, the onboard deliberative control system, on MBARI's CTD AUV platform was coupled with our supervised cluster learning algorithm. Our objective in this field trial was focused on the Elkhorn Slough which has been shown to have a strong plume signature during ebb tides. We had two objectives :
- Being able to adjust the spatial resolution of the volume survey based on the integrated plume signal strength over a transect.
- Firing the Gulper based on an instantaneous measurement of plume strength.
This was an
initial demonstration of T-REX adapting its generated plan based on
feedback from science instruments. To do so we connected T-REX to our
classifier to extract an
abstract view of the data which generated a
probability estimate of being in the plume. The classifer was trained
on data from
six previous missions. T-REX used this information to decide when to
fire the gulper (p(plume)>0.15) and decide the next waypoint in the
volume to adapt its spatial resolution. Transect separation within a
lawnmover patttern would change from 400m to 100m based on the
probability of having seen a plume during the previous transect.
The data below summarizes the two most representative runs on this day. We started with a test run similar to the actual science mission to demonstrate on a very small area with no Gulper firing for an initial validation followed subsequently by the actual volume survey in the red box on figure on the right.
Our results for this day are summarized as follows.
Science mission :
Our
survey started late at the tail end of the ebb tide on the South-West
node (D). The input specification was driven by a high-level goal
driven by the following key parameters:
- 4 points of a trapezoid to define the surface area
- min and max depth to bound the volume vertically
- resolution ranges: low, medium and high
- signal thresholds: firing the gulper, adjusting spatial resolution
- max 'gulps' per traverse
- minimum spatial separation between 'gulps'
TREX then plans and executes the mission, adjusting as it goes for
navigation and instrument control.
While T-REX was only
able to do perform two transects before terminating the mission on the
surface, due to a modeling error, the classifier nevertheless
identified a portion of the water during the second of two
transects as a plume resulting in two Gulper firings. In addition to
the adaptivity demonstrated by the classifier, the T-REX model also
encapsulated
spatial constraints (not
to fire less than 300m apart) in addition to not triggering more than
two
Gulpers on one transect.
The following figures show the track lines made by the AUV during this mission, the CPU usage of T-REX during this mission, the processed data from this mission and a view of the probability to be in a plume as seen by T-REX.
This last figure also shows the locations where the Gulper samples were taken (diamonds) which are in the area were T-REX considered the probability of being in the plume as the strongest.
Mission path(2D) 
Mission path(3D)
T-REX
CPU usage 
Processed science
data
Plume
mapping : classification result and gulper firing. Post-mission analysis revealed that there was an apparent run-time error
in the calibration of the CTD that was being used for
tracking salinity in-situ. This resulted in the CTD indicating a shift
in its salinity measurements resulting in a strong probability estimate
generated by the classifier incorrectly. However, T-REX's response to
this data was still appropriate to what sensory information was
presented since the planner adapted to the correct stimulus. Thus the
result was a validation of the system despite the evidence that a plume
did not exist in the patch of water visited in the second transect.
We were thus able to demonstrate
that sensor data was able to influence deliberation which in turn
influenced the navigational track of the vehicle while firing the
Gulper at appropriate locations. Lab analysis of the
Gulper samples are still forthcoming. The principal success of this
cruise was in demonstrating the tight integration between learning
and in-situ deliberation by showing a seamless integration of direct
sensor data and derived sensor data from a classifier to impact
navigation and science instrument control.