Adaptive Autonomous Robot Teams for Situational Awareness Georgia Tech’s Role University of

advertisement
Adaptive Autonomous Robot Teams
for Situational Awareness
Georgia Tech’s Role
University of
Pennsylvania
GRASP
Personnel
Georgia Tech

Faculty
Prof. Ron Arkin
 Prof. Tucker Balch
 Dr. Robert Burridge


GRAs
Keith O’Hara
 Patrick Ulam
 Alan Wagner
 Matt Powers

Mobile Intelligence Inc.

Dr. Doug MacKenzie
University of
Pennsylvania
GRASP
Impact – GT Role
• Provide communication-sensitive planning and behavioral
control algorithms in support of network-centric warfare, that
employ valid communications models provided by BBN
• Provide an integrated mission specification system
(MissionLab) spanning heterogeneous teams of UAVs and
UGVs
• Demonstrate warfighter-oriented tools in three contexts:
simulation, laboratory robots, and in the field
University of
Pennsylvania
GRASP
Communication Sensitive Planning
• Provide support for terrain models and other
communications relevant topographic features to
MissionLab
• Use plans-as-resources as a basis for multiagent
robotic communication control (spatial, behavioral,
formations, etc.) and integrate within MissionLab
University of
Pennsylvania
GRASP
Plans as Resources
• Motivated by Payton’s work.
• A precompiled map is an enabling resource.
• Maps converted to a two dimensional gradient
mesh a priori using A*.
• Robot queries “internalized plan” for directional
“advice” in the form of a vector.
• Queries and advice production are near real-time.
University of
Pennsylvania
GRASP
Internalized Plan as Behavior
• The GoToMapVector assemblage
controls retrieval of plan vectors from
maps, and consists of the following subassemblages:




University of
Pennsylvania
GetMapVector: Retrieves and injects a map
vector
Wander: Inject noise
Avoid Obstacles
MoveToGoal: Only used in experiments of
mixed reactive/planning behavior.
GRASP
Parallel Internalized Plans
• Different internalized
plans can be
combined by fusing
individual plans.
• Base plan contains
only physical objects.
• Other plans contain
additional constraints.
• The robot queries
advice from the most
constrained plan
(pessimistic).
University of
Pennsylvania
GRASP
Serial Internalized Plans
• Different internalized
plans are used one
after another.
• Each plan offers
situation specific
advice.
• Perceptual triggers
transition from only
plan to another.
• Opportunity for
contingency plans.
University of
Pennsylvania
GRASP
University of
Pennsylvania
GRASP
Initial Results
• Additional resources in the
form of internalized plans
aids team communication.
120
Percent Mission Time
• No difference results when
using reactive behaviors
vs. communication
insensitive plans.
Percent Time as One Network
100
80
60
40
20
0
• Communication planning
in serial and parallel result
in significant improvement
in communication.
University of
Pennsylvania
GRASP
Reactive
No
Comm.
Comm.
Comm. Planning Planning
Planning Parallel
Serial
Experiment
Plans as Resources:
Upcoming work
• Conduct tests on teams of real robots.
• Determine the systems localization and map
accuracy requirements.
• Develop techniques for dealing with localization
errors and map inaccuracies.
• Extend the planning to 3D and generalize to other
space-time dimensions for multi-robot coordination
University of
Pennsylvania
GRASP
Communication-sensitive Team Behaviors
• Generation and testing of a new set of
reactive communications preserving and
recovery behaviors
• Creation of communications recovery and
preserving behaviors sensitive to QoS
• Expansion of behaviors in support of lineof-sight and subterranean operations
University of
Pennsylvania
GRASP
Communications Recovery Behaviors
• Retrotraverse: Log robot’s position at regular intervals;
when comms breaks, move to last N positions logged until
comms recovered
• Move to Higher Ground: Use inclinometer data to guide
ascent to vantage point for communications recovery
• Nearest Neighbor: Track the last known position of
connected robots; if comms lost, move towards the nearest
robot’s last position
• Bridging: Couple separated networks by tracking positions
and moving towards location of network lesion; currently
UAV behavior
• Shepherding: Search out robots that have been cut off
from the network; once found, guide back (currently UAV)
University of
Pennsylvania
GRASP
University of
Pennsylvania
GRASP
Experimental Design

Missions run on simulated Quantico map

20 trials starting at regularly spaced intervals along the
western side of the map and moving to a central
location on the eastern side of the map

2 UGVs moving in a line formation with 20m spacing

Recovery behaviors used in isolation of one another

Metrics: Mission Completion Rate, Recovery time
University of
Pennsylvania
GRASP
Results
Trials Completed
Number of Trials Completed
20
18
16
14
12
10
8
6
4
2
0
19
18
No Preserving, No Recovery
No Preserving, Higher Ground
No Preserving, Nearest Neighbor
11
No Preserving, Retrotraverse
8
Maintain Signal Strength, No Recovery
Maintain Signal Strength, Higher Ground
2
0
0
0
Maintain Signal Strength, Retrotraverse
Maintain Signal Strength, Nearesr Neighbor
Communication Sensitive Behaviors
Using the Nearest Neighbor Recovery behavior approximately 50% of the
trials were finished completely autonomously
Retrotraverse and Move to Higher Ground were usually not able to finish
the trials autonomously by themselves and will require transitions/planning
once communications recovered
University of
Pennsylvania
GRASP
Results (2)
Communications Recovery Time
600
500
No Preservat ion, No Recovery
No Preservat ion, Higher Ground
Time
400
No Preservat ion, Ret rot raverse
No Preservat ion, Nearest Neighbor
300
M aint ain Signal St rengt h, No Recovery
M aint ain Signal St rengt h, Higher Ground
200
M aint ain Signal St rengt h, Ret ot raverse
M aint ain Signal St rengt h, Nearest Neighbor
100
0
Communications Behaviors
Retrotraverse results in the most rapid communications recovery of the
behaviors tested.
Move to higher ground results in the slowest recovery rate, largely due to
failure when the terrain was level.
Nearest Neighbor was successful in most cases, except in some situations
around buildings where the attraction to the lost robot and the repulsion to
the building that severed communications causes a local minima
University of
Pennsylvania
GRASP
Summary: Communications Recovery
• Retrotraverse provides the most rapid
communications recovery

Retrotraverse must be augmented with supplementary
behaviors or teleoperation to complete mission
• Move to Higher Ground and Nearest Neighbor
perform effectively in many cases


There are a number of cases where the behavior will
perform suboptimally
Supplementary behaviors or a more complex behavioral
selection may further improve results
University of
Pennsylvania
GRASP
Future Work
• Investigate means in which to activate recovery
behaviors based on available perceptual features
• Integration of cognizant failure (Gat) for recovery
behaviors
• Evaluate performance of recovery behaviors in the
context of larger teams, increased formation size,
and disparate goals
University of
Pennsylvania
GRASP
Communication-Preserving Behaviors
with Limited Memory
Value-Based One-Step Look-Ahead

Uses predictions of communication quality short
distances from current position to “hill-climb” to
better locations with respect to communication

Currently assumes teammates remain still when
predicting communication quality to reduce
complexity
University of
Pennsylvania
GRASP
Communication-Preserving Behaviors
• Operation:

Predict communication quality at locations a
small distance away using
 Map
information
 Network attenuation model
 Teammates assumed to remain still

Create motion vector based on predicted and
current communication quality
 Bearing
based on predicted quality
 Magnitude based on current quality
University of
Pennsylvania
GRASP
Communication-Preserving Behaviors
X
Predicted
communication
qualities
(r = .89)
Resulting vector
X
(r = .74)
Current
communication
quality
X
(r = .70)
(r = .68)
X
University of
Pennsylvania
GRASP
(r = .85)
Communication-Preserving Behaviors
Without Look-Ahead Behavior:
Obstacle-splitting
endangers
communication
quality
University of
Pennsylvania
GRASP
Communication-Preserving Behaviors
With Look-Ahead Behavior:
Obstacle-splitting
phenomena
eliminated
University of
Pennsylvania
GRASP
Communication-Preserving Behaviors –
1 step
• Future work:

Extend behavior to larger groups

Perform quantitative tests
 Compare
to other communication-preserving
behaviors
 Identify situations where most effective

Integrate into larger scenarios
University of
Pennsylvania
GRASP
Memoryless Communication Preserving Behavior
Maintain-Signal-Strength
• Servos on signal strength to preserve communication.
• Sum over every “connected” robot


Vector_Magnitude = (T-R)/T when (T-R) > D
Vector_Direction = angle to the robot
where T: Target Signal Strength, D: Signal Deadzone, R: Actual
Signal strength
• Connected can be defined to mean either directly
connected or connected via a multi-hop route.
University of
Pennsylvania
GRASP
Illustration of Maintain-Signal-Strength
g1
g2
Communication
Quality Increases
Communication
Quality Decreases
s1
University of
Pennsylvania
GRASP
s2
University of
Pennsylvania
GRASP
Communication Preservation
Experiments
•
•
•
•
Mission: Each robot navigates to its goal.
Team Sizes: 2, 4, 6, and 8
Distance separating robots: 10, 20, 40 meters
25 random worlds
 12% obstacle coverage
 256 x 256 meters
• Three behaviors are compared.
 No communication behavior (control)
 MSS using positions of directly connected robots
(single-hop)
 MSS using all available positions (multi-hop)
University of
Pennsylvania
GRASP
Percentage of Time as One Network
• Some communication strategy is
needed to keep the network one as
you increase the distances or the
number of robots.
•There doesn’t seem to be a
significant difference between the
two variations of the behavior.
University of
Pennsylvania
GRASP
Mission Completion Time
• Both variations of the behavior add
a significant amount of time to
mission completion.
University of
Pennsylvania
GRASP
Communication Models and Fidelity
• Working with BBN to incorporate suitable
communication models into MissionLab in support
of both simulation and field tests
University of
Pennsylvania
GRASP
Current Network Model Status
• Models wireless communication networks in
3 dimensions.
• Integrated into MissionLab
• Signal Attenuation

Free-space path-loss


Dependent on distance between robots, frequency of
communication band, and antennae height.
Line-of-Sight Obstructions
Absolute signal attenuation.
 Obstructions modeled as arbitrary polygons or right
cylinders with height.
 Terrain map can be used which can occlude LOS.

University of
Pennsylvania
GRASP
The Quantico Overlay From a
Communications Perspective
University of
Pennsylvania
GRASP
Next Steps in Modeling Network
• Obstructions will attenuate signal at different
magnitudes.

Model buildings and foliage.
• Accurate model of signal attenuation over
rough terrain.
• Mimic capabilities of BBN “black-box”
• Understand how different levels of model
fidelity impact multi-robot team performance.
University of
Pennsylvania
GRASP
Communication-sensitive Mission Specification
• MissionLab is a usability-tested Mission-specification
software developed under extensive DARPA funding
(RTPC / UGV Demo II / TMR / UGCV / MARS / FCS-C programs)
• Using MissionLab as a basis:



University of
Pennsylvania
Adapt to incorporate air-ground communicationsensitive command and control mechanisms
Extend to support physical and simulated experiments
for objective air and ground platforms
Incorporate new communication tasks and triggers
GRASP
MissionLab’s Spatial Planner
• Incorporates Navigator Component of the AuRA architecture
- A map of obstacles is read in by the system
- The map is “grown” to represent configuration space
- The free space is partitioned into a collection of convex
“meadows”
- Start and End points are selected by the user
- The planner performs A* search to find an initial path
- The path is improved by tautening
• Can be invoked from MissionLab’s cfgedit tool
• Creates an FSA series of waypoints
University of
Pennsylvania
GRASP
Initial Map and Meadow Map
University of
Pennsylvania
GRASP
Path Chosen and Formation Run
University of
Pennsylvania
GRASP
Technology Integration
• Conduct Early-on
Demonstrations on
Ground Robots at GT
• Provide our Hummer
Command and Control
Vehicle for Team support
at Objective
Demonstration
University of
Pennsylvania
GRASP
Interface Control Document
• To explicitly capture all aspects of all
interconnections between project components.



Communications protocols, frequencies, and timing
Language and data formats
Experimental communications fault injection
• To define new mission description language:
CMDL+
• To detail communications-sensitive behaviors
developed by project teams.


Communication-preserving
Communication-recovering
University of
Pennsylvania
GRASP
(Mounted in
GT Hummer)
PEN
N
ROCI
GaTec
h
MLab
VIP
Displa
y
USC
USC
Helo
Playe
r
University of
Pennsylvania
GRASP
BBN
PENN
GPS Jammer
• Supports evaluation of robot localization
methods in challenging environments
• White noise centered on selected frequency
• Power: 50 to 200mw (about 50-100 meters)
• Performance to be characterized in the
coming few weeks
Engineered by Daniel Walker (BORG Lab)
University of
Pennsylvania
GRASP
Summary - Georgia Tech Contributions
• Communications Sensitive Behaviors


Preserving
Recovery
• Communications Planning Behaviors



Plans as Resources
One-step planning
Team spatial waypoint planning
• Infrastructure





Communications models support
MissionLab as an integration vehicle
ICD Development lead
Hummer base station / Test equipment
Scenario development
University of
Pennsylvania
GRASP
Backup Slides
University of
Pennsylvania
GRASP
Plans in Serial Demo explained
Seven plans are used in this demo
University of
Pennsylvania
GRASP
Download