Coordinating a Fleet of Autonomous Underwater Gliders Using a DecisionTheoretic Approach in a Multia

1
Coordinating a Fleet of Autonomous Underwater
Gliders Using a Decision-Theoretic Approach in a
Multi-agent System
Chhaya Mudgal1, Scott Glenn1, Oscar Schofield1,
Clayton Jones2, Douglas Webb2, Gary
Kirkpatrick3 1Rutgers University, New Brunswick,
NJ, 2Webb Research Corp, Falmouth, MA, 3Mote
Marine Lab, Sarasota, FL
Glider Mission Control Center
Approach
Sample Products
Introduction
Data Sets
Agent Oriented Software
I walk into our control room, with its panoply
of views of the sea. There are the updated global
pictures from the remote sensors on satellites,
there the evolving maps of subsurface variables,
there the charts that show the position and
status of all our Slocum scientific platforms,
and I am satisfied that we are looking at the
ocean more intensely and more deeply than anyone
anywhere else.
- Henry Stommel,
The SLOCUM Mission

• This research focuses on the design and
  implementation of a centralized control system to
  coordinate a fleet of Gliders based on data
  available from the fleet or other scientific
  systems. The Glider Mission Control Center is
• Autonomous
• Adaptable
• Adaptive
• Responsive
• Flexible

Figure 9 is a sample plot produced by the Glider
Data Agent using NOAAs SGT showing a temperature
profile collected by a Glider. The Glider Data
Agent finds the location of the maximum
temperature gradient (green star) and passes this
thermocline location on to the Mission Control
Agent. The Mission Control Agent uses this
information to determine the vertical extent of
the undulations. Similar plots are implemented
for conductivity,salinity and density.
The Slocum Glider tested at LEO was equipped with
a SeaBird CTD. Numerous navigation and control
tests were conducted during 2000, including
station keeping in strong currents, steering a
course with and without current corrections,
conducting a cross-shelf survey transect, and
monitoring the location of a front overnight
until a boat could return the next day. Before
the Glider was recovered, three different
shipboard CTDs were profiled as the Glider
undulated nearby.

• Implementation of the Glider Mission Control
  Center is based on Agent Oriented programming.
• What are Software Agents?
• A Software Agent is a concept from Artificial
  Intelligence. Software agents are
  computational entities which are capable of
  working autonomously in environments inhabited
  by other agents.
• Software agents can
• perceive their environment
• act upon the environment
• communicate with other agents

What is a Slocum Glider? The Slocum Glider is an
Autonomous Underwater Vehicle (AUV) that
undulates in the water column by changing its
buoyancy. The Coastal Electric version changes
its buoyancy using an electric hydraulic pump to
inflate and deflate an internal diaphragm. Wings
translate the vertical motion into a forward
velocity, resulting in a sawtooth sampling
pattern along a subsurface transect. The vehicle
is steered to a specific location, either by a
rudder (for fast response), or by lateral changes
in its center of gravity (similar to a
hang-glider). The Slocum Glider is equipped with
a payload bay that nominally contains a CTD. At
regular intervals, the Glider surfaces to obtain
another GPS fix and communicate with a shorebase
using the antennas located in the tail fin.
Sensor, vehicle status, and position data can be
transmitted to shore, and new missions can be
downloaded. The Slocum Glider is named after
Joshua Slocum, the first person to sail solo
around the world.
Mission Control Center is a Multi-Agent Decision
Making System consisting of following Agents
Glider Data Agents A new Glider Data Agent
thread is spawned for every Glider deployed. Each
Glider Data Agent monitors their specific Glider
for the availability of new data. When new data
is acquired, it is processed, analyzed, and
graphically displayed. Results are communicated
to the Mission Control Agent. Mission Control
Agent The Mission Control Agent is the central
decision making entity of the Control Center.
After receiving data products and status
information, initial tasks are to update a status
board locally and on the World Wide Web, and to
send email warnings of potential problems to
controllers. It then evaluates the data products
to decide whether to continue the present mission
or to design and start a new mission.
Figure 3 Comparison of Glider and three
different shipboard CTD profiles
Figure 3 shows the temperature and salinity
profile comparison, and Figure 4 illustrates the
temperature section from the cross-shelf
transect.The LEO Glider missions were conducted
in the vicinity of a cross-shelf array of
bottom-mounted ADCPs and ADPs deployed for model
validation. Whenever the Glider was within two
kilometers of one of the validation array
moorings, the Glider pressure record was used to
average the observed velocities. Figure 5
compares the along-shelf and cross-shelf
components of the depth average velocity computed
from the current meters and the Gliders.
Figure 9 Glider Temperature Profile
Figure 10 is an example produced by the Mission
Control Agent using Open Map to track the status
of a hypothetical fleet of 8 Gliders conducting a
coordinated cross-shelf survey of the New Jersey
Shelf. The intended flight lines are shown, with
the current location of each Glider marked by the
green arrow.
Figure 6 Abstract Agent Model
Who uses Agent technology? Software Agents are
being used to construct self-aware,
self-controlled and self-operated robots,
exploring rovers and intelligent machines. The
most celebrated use of software agents is by
NASAs Deep Space 1 spacecraft. Deep Space 1s
Remote Agent is the first artificial
intelligence software to command a spacecraft.
Remote Agent detected, diagnosed and fixed
problems, making decisions based on high level
goals to keep the mission on track. Remote
Agents most recent achievement was the
successful navigation of the spacecraft through
the tail of comet Borrelly.It is envisioned that
spacecraft equipped with agent oriented software
will enable NASA to pursue missions that once
were considered too elaborate, too costly, or too
dependent on large teams of ground controllers.
Slocum specifications Weight
56 Kg Hull Diameter 21.3 cm Vehicle Length
1.5 m Depth Range 4 - 200
m Energy Alkaline Batteries Endurance
30 days Range 1500
km Navigation GPS and internal
dead reckoning,
altimeter Horizontal Speed
35cm/sec (30km/day)
Figure 5 ADCP versus Glider Drift Comparison
Figure 4 Temperature section July 19, 2000
Communication
Figure 10 Map showing Glider Status
Ongoing Work
Webb Research Corporation and the University of
Washington/APL are collaborating on the
development of Iridium satellite phone modems to
provide Autonomous Underwater Gliders the global
communication capabilities required for long-term
deployments. Glider communications currently
rely on range-limited cell phone or RF modems. At
LEO, the RF modem repeater located on the 64 m
Tuckerton meteorological tower provides a 30 km x
30 km operating area centered on the
high-resolution CODAR surface current field.Tests
are normally conducted in the middle of this
field, so if a Glider is lost at night, it is
unlikely to drift out of range before daybreak.
For backup, a second RF repeater was installed on
the tail of the PHILLS survey aircraft also
operating at LEO. During the 2001 steering tests

Figure 1 Slocum Glider
1) Continue to expand the functionality of the
mission control center a) Vertical control b)
Horizontal control c) Coordinated fleet
operations 2) Improve satellite communications 3)
Addition of the Hydroscat-2 Fluorometer/Backscatte
r sensor for phytoplankton and particle
concentration. 4) With Mote Marine Lab,
miniaturization and installation of a
spectrophotometer to measure absorption
signatures that delineate phytoplankton load
and composition. To be tested for Red Tide
detection and tracking.
An academic/industry partnership between Rutgers
University and Webb Research Corporation resulted
in a series of test flights of the Slocum Coastal
Electric Glider at Rutgers' Longterm Ecosystem
Observatory (LEO) during the 1999, 2000 and 2001
summer Coastal Predictive Skill Experiments. The
first open ocean tests of the Coastal Electric
Glider were conducted with a safety tether in
1999 and unteathered in 2000. A new version with
faster steering response was tested in 2001in the
steep ridge and swale topography to the south of
LEO.
over the steep topography, the Glider stopped
reporting in at its scheduled intervals. A boat
search of the vicinity guided by the CODAR
current fields revealed nothing, leading us to
believe that the Glider had grounded. In this
situation, the Glider drops a ballast weight
after 12 hours and returns to the surface. The
expected release time was during a scheduled
survey mission of the PHILLS aircraft. A boat
was sent to the last known location. At the
expected release time the Glider surfaced and
transmitted
Figure Deep Space 1 flyby of comet Borrelly
Inter-Agent Communication
As Software Agent technology becomes more
prevalent, the demand for efficient and versatile
agents will grow. Designing a single agent to
perform all the tasks increases the complexity of
development. An alternative is to design a
society of agents with each agent handling its
own task efficiently. For the society of agents
to be able to function coherently, cooperation
and collaboration is an important requirement.
Inter-agent cooperation and collaboration is
highly dependent on a common communication medium
and a common communication language. There are
many agent frameworks, languages and protocols
available. The software package for our agent
implementation Java Agent Toolkit (implemented
at Stanford University) uses KQML Knowledge
Query Manipulation Language for inter-agent
communication.
RF Repeater
its location through the repeater on the aircraft
directly to the vessel at sea.This rapid rescue
is one demonstration of the utility of long-range
communication systems.
Decision Making
Motivation

• Decision making methodology must be incorporated
  in the semi-autonomous software tool that
  coordinates the fleet of underwater Gliders. The
  Decision Theoretic approach will be used to
  implement decision making in our system.Decision
  Theory is a mathematical framework for
  determining the best action given a decision
  problem. It is a process of sequential decision
  making that provides structure to a complex
  problem, identifies important objectives and
  generates alternative courses of action.In our
  context, decision making is data, knowledge and
  communication driven.
• Whenever there is incomplete knowledge or no
  knowledge, it creates a situation of
  uncertainty. For example, setting the depth of
  the Glider undulations based on the thermocline
  must take into consideration the possibility that
  there is no thermocline or the thermocline is not
  distinct.The Decision Theoretic tool adopted here
  is an Influence Diagram. Influence Diagrams are
  an extension of Bayesian Belief Networks
  containing decision and utility variables.
  Actions are selected by evaluating the decision
  network for possible choices.

Figure 8 Flowchart diagram for the Mission
Control Center
LEO Glider Deployments
Implementation
With global satellite communication
capabilities, Autonomous Underwater Gliders will
have the ability to patrol the subsurface ocean
for long durations with reduced fear of being
lost at sea. A small fleet of Gliders flying
beneath the satellite- and CODAR-derived surface
fields will provide 3-d information for
assimilation in forecast models.Full water column
undulations will provide data on the temperature
and salinity structure below the satellite SSTs.
Undulations above or below the thermocline will
provide depth average current estimates to
improve the assimilation of CODAR surface
current maps

• The software implementation for graphics and data
  analysis uses Suns Java Development Toolkit 1.3
  (JDK 1.3)
• Software packages used for the graphical
  representation and Agent Implementation are
• NOAAs-Scientific Graphics Toolkit (SGT). SGT is
  a Java library package that provides platform
  independent and highly interactive graphics for
  scientific data.
  
  
• BBN Technologies-OpenMap. OpenMap is a package
  based on JavaBeans. Its library package allows
  one to view and manipulate geospatial data.
• Stanford Universitys-Java Agent Template (JAT
  Lite). JAT Lite is a package of programs written
  in Java that allows one to create software
  agents. These software agents can send and
  receive messages using the standard agent
  communication language KQML. JAT Lite has been
  used to implement the Glider Control Agent and
  the Mission Control Agent.

The bathymetry map at left illustrates the Glider
GPS track for the 2000 (white) and 2001 (red)
deployments. Two-way communications between the
Glider and the Tuckerton field station were
achieved using a radio modem repeater located on
the 64 m high meteorological tower.
Acknowledgements
Figure 7 Types of Decision Making
The Rutgers/Webb partnership was initiated by
NOPP in 1999 and is currently maintained by ONR
and the Great State of New Jersey. The
partnership was expanded to include Mote Marine
Lab in 2002 by NSF. Additional thanks to Dr.
David Fratantoni (WHOI) and his group for their
contributions to the summer 2000 Glider
deployment at Tuckerton. This poster can be
viewed online http//marine.rutgers.edu/Cool/CoolR
esults/agu2002.
Figure 2 Glider maps from 2000 and 2001