Coverage of a given area by means of coordinated autonomous robots is a mission
required in several applications such as, for example, patrolling, monitoring or
environmental sampling. From a mathematical perspective, this can often be
modeled as the need to estimate a scalar field, eventually time varying as in
the security applications. In this paper, the problem is addressed for the
challenging underwater scenario, where localization and communication pose
additional constraints. The solution exploits the appealing properties of the
Voronoi partition of a convex set within a probabilistic framework. In addition,
the algorithm is totally distributed and characterized by a strong engineering
perspective allowing the handling of asynchronous communication or possible loss
or adjunct of vehicles. Beyond the test in dozen of numerical case studies, the
algorithm has been validated by a challenging underwater test in 3 dimension
involving two Autonomous Underwater Vehicles (AUVs). The experiments were run in
the La Spezia harbor, in Italy, in February 2012 as demo
of the European project \co3auvs.
1. Experimental Results of Coordinated Coverage by
Autonomous Underwater Vehicles
Alessandro Marino, Gianluca Antonelli
Universit`a di Salerno, Italy
Universit`a di Cassino & ISME (Integrated Systems for Marine Environment), Italy
antonelli@unicas.it
http://webuser.unicas.it/lai/robotica
http://www.isme.unige.it/
Marino, Antonelli Karlsruhe, 9 May 2013
2. CO3
AUVs
Cooperative Cognitive Control of Autonomous Underwater Vehicles
fundings : European FP7, Cognitive Systems, Interaction, Robotics
kind : Collaborative Project (STREP)
duration : 3 years, 2009-2012
partners : Jacobs University, DE;
ISME, I;
Instituto Superior T´ecnico, P;
GraalTech, I
http://www.Co3-AUVs.eu
Marino, Antonelli Karlsruhe, 9 May 2013
3. Problem formulation
Multi-robot harbor patrolling
Totally decentralized
Robust to a wide range of failures
communications
vehicle loss
vehicle still
Flexible/scalable to the number of vehicles add vehicles anytime
Possibility to tailor wrt communication capabilities
Not optimal but benchmarking required
Anonymity
To be implemented on a real set-up obstacles. . .
Marino, Antonelli Karlsruhe, 9 May 2013
4. Proposed solution
Proper merge of the Voronoi and Gaussian processes concepts
Motion computed to increase information
Framework to handle
Spatial variability regions with different interest
Time-dependency forgetting factor
Asynchronous spot visiting demand
Mathematically strong overlap with (time varying) coverage,
deployment, resource allocation, sampling, exploration, monitoring, etc.
slight differences depending on assumptions and objective functions
Marino, Antonelli Karlsruhe, 9 May 2013
5. Proposed solution
Proper merge of the Voronoi and Gaussian processes concepts
Motion computed to increase information
Framework to handle
Spatial variability regions with different interest
Time-dependency forgetting factor
Asynchronous spot visiting demand
Mathematically strong overlap with (time varying) coverage,
deployment, resource allocation, sampling, exploration, monitoring, etc.
slight differences depending on assumptions and objective functions
Marino, Antonelli Karlsruhe, 9 May 2013
6. Background
theoretical details
Antonelli, Chiaverini, Marino, A coordination strategy for multi-robot
sampling of dynamic fields, ICRA 2012
experimental validation with surface vehicles
Marino, Antonelli, Aguiar, Pascoal, Multi-robot harbor patrolling: a
probabilistic approach, IROS 2012
Marino, Antonelli Karlsruhe, 9 May 2013
7. Voronoi partitions I
Voronoi partitions (tessellations/diagrams)
Subdivisions of a set S characterized by a metric with respect to a
finite number of points belonging to the set
union of the cells gives back the set
the intersection of the cells is null
computation of the cells is a
decentralized algorithm without
communication needed
Marino, Antonelli Karlsruhe, 9 May 2013
9. Background I
Variable of interest is a Gaussian process
how much do I trust that
a given point is safe?
Given the points of measurements done. . .
Sa = (xa
1 , ta
1 ), (xa
2 , ta
2 ), . . . , (xa
na
, ta
na
)
and one to do. . .
Sp = (xp, t)
Synthetic Gaussian representation of the condition distribution
ˆµ = µ(xp, t) + c(xp, t)TΣ−1
Sa(ya − µa)
ˆσ = K(f(xp, t), f(xp, t)) − c(xp, t)TΣ−1
Sac(xp, t)
c represents the covariances of the acquired points vis new one
Marino, Antonelli Karlsruhe, 9 May 2013
10. Description I
The variable to be sampled is a confidence map
Reducing the uncertainty means increasing the highlighted term
ˆµ = µ(xp, t) + c(xp, t)TΣ−1
Sa(ya − µa)
ˆσ = K(f(xp, t), f(xp, t)) − c(xp, t)T
Σ−1
Sac(xp, t)
ξ
− > ξ example
Marino, Antonelli Karlsruhe, 9 May 2013
11. Description II
Distribute the computation among the vehicles
each vehicle in its own Voronoi cell
Compute the optimal motion to reduce uncertainty
Several choices possible:
minimum, minimum over an
integrated path, etc.
Marino, Antonelli Karlsruhe, 9 May 2013
14. Accuracy: example
Only the restriction to V or2 is needed for its movement computation
τs
x1 x2 x3 x4
x
ξ(x)
V or2
Marino, Antonelli Karlsruhe, 9 May 2013
15. Accuracy: example
Merging of all the local restrictions leads to a reasonable approximation
τs
x1 x2 x3 x4
x
ξ(x)
V or2
Marino, Antonelli Karlsruhe, 9 May 2013
17. Numerical validation
Dozens of numerical simulations by changing the key parameters:
vehicles number
faults
obstacles
sensor noise
area shape/dimension
comm. bit-rate
space scale
time scale
2
3 4
Marino, Antonelli Karlsruhe, 9 May 2013
18. Some benchmarking
With a static field the coverage index always tends to one
0 200 400 600 800 1000
0.2
0.4
0.6
0.8
1
step
[]
Coverage Index
Marino, Antonelli Karlsruhe, 9 May 2013
19. Some benchmarking
Comparison between different approaches
00
Lawnmower
Proposed
Random
Deployment0.5
1.5
2
200 400 600 800 1000 1200
1
[]
step
same parameters
lawnmower rigid wrt
vehicle loss
deployment suffers
from theoretical
flaws
Marino, Antonelli Karlsruhe, 9 May 2013
20. Vehicle characteristics
internal diameter .125 m
external diameter .14 m
length 2 m
mass 30 kg
mass variation range .5 kg
(at water density 1.031 kg/m3
)
moving mass max displacement 0.050 m
Lead acid batteries 12 V 72 Ah
autonomy at full propulsion 8 h
diving scope 0–50 m
break point in depth 100 m
speed with jet pump propeller 1.01 m/s 2 knots
speed with blade propeller 2.02 m/s 4 knots
cpu 1GHz, VIA EDEN
dram 1GB, DDR2
Marino, Antonelli Karlsruhe, 9 May 2013
21. Experimental validation
joint experiment with Graaltech NURC (NATO Undersea Research
Center) facilities, La Spezia, Italy
Marino, Antonelli Karlsruhe, 9 May 2013
22. Experimental validation
2 F`olaga, 4 acoustic transponders, 1 gateway buoy
110 × 80 × 4 m
1.5 m/s
33 minutes
WHOI micromodem 80 bps
Time Division Multiple Access
localization: every 8 s
user comm: 31 byte/min with 14 s delay
Marino, Antonelli Karlsruhe, 9 May 2013
23. Experimental validation
Due to poor communication, the algorithm runs by predicting the
movement of the other
# fields size (bytes)
1) vehicle ID 2
2) localization time 4
3) vehicle latitude 4
4) vehicle longitude 4
5) vehicle depth 4
6) target latitude 4
7) target longitude 4
8) target depth 4
Marino, Antonelli Karlsruhe, 9 May 2013
24. Experimental validation - video
Coverage index
200 400 600 800 1000 1200 1400 1600
0.1
0.2
0.3
0.4
[]
0.5
00
time [s] 1800
Marino, Antonelli Karlsruhe, 9 May 2013