cooperative auv 2

8/2/2019 Cooperative Auv 2

1/39

localization and environmentallocalization and environmental

sampling by AUV teamssampling by AUV teamsAndrea Caiti (1,2), Pino Casalino (1,3), Andrea Munaf (1,2),

Enrico Simetti (1,3), Alessio Turetta (1,3)

Andrea Caiti (1,2)

, Pino Casalino (1,3), Andrea Munaf (1,2),

Enrico Simetti (1,3), Alessio Turetta (1,3)

(1) ISME Interuniv. Res. Ctr. Integrated Sys. Marine Env.

(2) DSEA & Centro Piaggio - University of Pisa, Italy

(1) ISME Interuniv. Res. Ctr. Integrated Sys. Marine Env.

(2) DSEA & Centro Piaggio - University of Pisa, Italy

,

[email protected]

,

[email protected]


2/39

MenuMenu

AUVs for environmental sampling

propelled

hybrid

Adaptive sampling and autonomydata-driven vs. model-driven

Limitations to uw robot team cooperationcommunication

localization

Adaptive sampling in a communication


3/39

-

replace ship and/or buoys with autonomous vehicles

pre-program the mission, receive data at a remote station

CTDs, water quality, bathymetry, seabed morphology

data quality enhanced by the use of an underwater platform

New measurements possibilities

gradient-following, feature mapping, synoptic measurements

exploit the on-board intelligence as data are gathered:

adaptivity

ex loit the availabilit of multi le vehicles:

team cooperation


4/39

Some AUVs Some AUVs ceanograp c g ers

very long endurance

inexpensive Hybrid vehiclesoceanograp c samp ng

pre-programmed missioncooperation (Leonard et al.)

inexpensiveoceanographic sampling

-Propelled AUVs

deep waterlong endurance

cooperation

very expensive

seabed mapping

pre-programmed mission

shallow watersome endurance

moderately expensive

oceanography, seabed mapping

pre-programmed, cooperation


5/39

and some ASV - and some ASV -

Autonomous Surface VehiclesAutonomous Surface Vehicles

OASIS (NOAA) - Delfim (IST)

Scout (MIT) - Charlie (CNR-

ISSIA)


6/39

Adaptation and autonomyAdaptation and autonomy

Sam le at oints/ aths that maximize some fi ure of

merit related to our knowledge of the sampled field

In the oceanographic context adaptation can be:

Model driven: use model rediction to determine next mission no autonomous planning needed (but autonomous navigation

required)

open loop - feedback

Data driven: use data to determine next way-point

autonomous lannin as well as autonomous navi ation

feedback only


7/39

Where/when is data drivenWhere/when is data driven

Whenever occurrence of anomalous

events as to e mon tore , etecte ,

classified

. . Fixed sensors in key positions

Periodic sampling around the field

Other instances: water quality in touristic

industry discharge

security: intrusion detection,

m ne or ea searc


8/39

Some formal frameworksSome formal frameworksfor data-driven adaptive samplingfor data-driven adaptive sampling

increase sample density when the measured quantity is rapidly

varying (either in time and/or in space)

estimate variation rate from past or neighborhood data

deterministic setting: smoothness

probabilistic setting: information gain

app cat on- epen ent we g ng

use tools from both approaches


9/39

Deterministic setting: smoothnessDeterministic setting: smoothness

A measure of the variability of the field

weighed sum of the time and spatial derivativesn

F t b D F c D F = +r , , , ,n

( ; )nS

F t kr

projection over a space of basis functions, ordered with

increasing variability ( ; ) ( ; ) ;i i i iiF t h t h F

= =r r p f2

ii nh k

=


10/39

Probabilistic setting: information gainProbabilistic setting: information gain

Minimize the expected uncertainty of the field

A priori mean and covariance of the field( ; ) { ( ; ; )}F t E F t =r r

m measurements at points (ri ;ti), estimation algorithm

( , '; , ') {[ ( ; ) ( ; )][ ( '; ') ( '; ')]}C t t E F t F t F t F t = r r r r r r

choose the m samples so to minimize the a posteriori

( )( ; ) ( ; ), ( ; ) ; 1,..., )i iF t T F t F t i m= =r r r

uncer a n y o e es ma e e

{[ ( ; ) ( ; )][ ( ; ) ( ; )]}J E F t F t F t F t d dt= r r r r r

(adapted from Leonard et al., 07)


11/39

Sensors

deployment

ommun cat on

& fusionSense

e erm ne

next position

0

centralized

;i ir

each

nodejointly

str ute

collaborative

( ; )i iF tr ( ; ) ( ( ; ) )1,..., 0,...

i iF t T F t i k

== =

r r ;i itr

(adapted from Huguenin & Rendas, 06)


12/39

Model-driven predictor/corrector methodsModel-driven predictor/corrector methods

Sensor sam linKalman-like

Process Modelposition

estimation

static

ada tiveProcess( / 1)

/ 1

k k

k k

x

C

sampling

Compare &

Estimate

predictioncorrection

Optimize expected

( / )k kC

information gain


13/39

Cooperation and coordination

ada ted from Whitcomb and Yuh 09

Cooperation and coordination

ada ted from Whitcomb and Yuh 09

Team of heterogeneous vehicles

Payoff ada tation

same/better performance

no single point weakness

sensor networks

ubiquitous computing

mo e agents, consensus, ...

Drawback:coo eration and coordination re uire inter-

vehicle communications


14/39

The acoustic communication channelThe acoustic communication channel Transmission loss

limitations on channel capacity / requirements on source

level

Multi-path structure causes symbolic interference

limitations on codin /decodin schemes

Acoustic channel predictions used to determine

relativedepths and maximumrange among Tx/Rx to

guarantee a given SNR


15/39

Predicting channel characteristicsPredicting channel characteristics

--

SNR(x,w) = SL - TL(x,w) - N(w) Numerical Code

Equations for channel characterization

environmental conditions

Includes ( in TL):

[Bandwidth]

intrinsic attenuation

wave interference patterns

apac y

[TX Power]

range independent

stationary sources

(Stojanovic, 07) (Caiti et al., 09)


16/39

Limitations & opportunities forLimitations & opportunities forcooperative robotscooperative robots

Network connectivity constraints limit therelative mobilit of the a ents

Mobility of the agents can be exploited todynamically maximize some figure of merit of the

communication channel


17/39

What happens with COTSWhat happens with COTS

COTS acoustic modems have fixed bit rate and source

level, with BER depending on the SNR

Bit rate: from 256 b/s to optimistic 15 kb/s

of the individual agents and on parsimonious comms

overhead


18/39

Team localizationTeam localization

A team of AUVs must be cheap!

Underwater navigation is a major source of vehicle cost inertial systems

Use the acoustic modems also to localize acoustically

measurements

Some vehicles at the surface eo-referenced relativelocalization

DGPS


19/39

Cooperative uw localizationCooperative uw localization

A hot research topics J. Leonard; Sukhatme; Chitre; Antonelli ...

Alternate between ranging and communication

, , ...

Use depth sensors to convert from 3D to 2D

Several issues: linear vs. nonlinear estimate

observability depending on relative vehicle motion

on-board correction for bended ray-paths

clock synchronization

Our approach: distributed EKF with delay (from team

theory work dating back to the 80s!)


20/39

Designing behavioursDesigning behaviours

Not observable Observable EKF Error

(Antonelli, Caiti et al., ICRA 10)


21/39

Real-time Ray-tracing (RT2)Real-time Ray-tracing (RT2)

Compensate ray bending through a look-up table

en ng ray error: ncreas ng mpor ance w range

In distributed localization range errors will accumulate

- .

PTT:Partial

Travelled

PTD:

Distances


22/39

Look-up table and on-line estimateLook-up table and on-line estimate

receiver depth

V1 1

V1 2

. .

.

V1 n

V V . . V2 1 2 2 . 2 n

. . . . . .

. .

.

. .

.em

itter

de

pth

Vn 1

Vn 2

. .

.Vn n

])(,),(),([ 21 mabababab kVkVkVV L=

k

rrrrRTkV ababababababababab ],,,,|,,,,[)(43214321 = (Casalino et al.,

Oceans 10)


23/39

Cooperative adaptive sampling withCooperative adaptive sampling with

- Each vehicle- Each vehicle

per orms s own

vertical sampling

per orms s own

vertical sampling

- All the vehicles

share information

- All the vehicles

share information

(acoustic)(acoustic)

plan the next move of the team in order to optimize:plan the next move of the team in order to optimize:

- sampling map quality (map res. below a given threshold)- overall area coverage measurements (cover the area by sampling

where needed)

- sampling map quality (map res. below a given threshold)- overall area coverage measurements (cover the area by sampling

where needed)

- while maintaining connectivity

range constraint among the vehicles, varying in space and time

- while maintaining connectivity

range constraint among the vehicles, varying in space and time


24/39

deterministic data-driven approach

deterministic data-driven approach

Each vehicle computes its next admissible range for its next

measurement (admissible exploring radius) on the basis of the

Each vehicle computes its next admissible range for its next

measurement (admissible exploring radius) on the basis of the

local smoothness of the environmental map -Local computationslocal smoothness of the environmental map -Local computations

( ) ( ; ) ( ; )kF t h t

=r r : RBF family( ) ( )( ; ) ( ; ) ( ; ) ( )k kI IF t F t F t G h

r r rh: fill distance

G: known function


25/39

Optimize area coverageOptimize area coverage

Local decision rule based: maximize distance rom reviousLocal decision rule based: maximize distance rom revious

samples and from next location of the other vehiclessamples and from next location of the other vehicles

Information exchange needed

Information exchange needed


26/39

Communication/range constraintCommunication/range constraint

Next sampling points must preserve connectivity of the

communication structure

Next sampling points must preserve connectivity of the

communication structure


27/39

Serial graph structure, fixed topologySerial graph structure, fixed topology

Use of

Dynamic CTransition Cost

Programming

j

Possible local moves Next

graph

)pp(C 1jj +,

amp ng c rc es

(independently locally evaluated)

Backward phase Forward phase

Distributed implementation of dynamic programming!


28/39

Articulated chain structure,

d i l

Articulated chain structure,

d i lMinimum

algorithm

current graphnext graph

Compare with

d.p. solution


29/39

Dynamic programming on test dataDynamic programming on test data

3.5

4

4.5

5

38.4

38.5

38.6salinity (ppm)

4

4.5

5

38.4

38.6salinty

(ppm)

1.5

2

2.5

3

y(km

)

37.9

38

38.1

38.2

38.3

2

2.5

3

3.5

y(km)

38

38.2

0 1 2 3 4 5

0

0.5

1

x (km)

37.6

37.7

37.8

0

0.5

1

1.5

37.6

37.8

0 1 2 3 4 5

x (km)

Test data,

courtesy

A. Alvarez,

Reconstructed

Via Cooperative Sampling

NURC an nterpo at on


30/39

Graph theory based cooperationGraph theory based cooperation

Sound optimality proof

s r u e mp emen a on o cen ra ze a gor ms

Heavy comms overload

increase

Possibility of a truly local rule-based algorithm?

Yes behaviours

Different approaches and implementations nosystematic design rule in general, some cases analyzed

Not always optimality guaranteed


31/39

An example of cooperative algorithmAn example of cooperative algorithm

in security applicationin security application Goals

Critical asset protection

Maintaining acoustic connectivity among the team

Each agent/node

Builds a local map of channel characteristics and comms

performance

Updates the map when new environmental measurements

Adapts its behaviour to tackle changes in the environment

u e- ase e av our an po en a e s

(Caiti et al., 2009/10)


32/39

AUVs e ui ed with: Acoustic modem max range:RC

Detection sonar max range:RD

cover with the sonarsminx a ix x the greatest areaaround the asset to

rotect2

, ,i ji j D Dx x R R i j

+

2, :ii j Ci j x x R

closest neighborhood information


33/39

--

Rule 1: ove toward the asset

Rule 2:Move away from your closest neighbor

Implemented through gradients of artificial potential

- ,

Vehicle course: vector sum of the two contributions


34/39

Ineherently robust to communication loss &

equipment failure!

Can include sonar/modem directionality


35/39


36/39

With omnidirectional sonar/modems infinite solution

exists (all symmetric configurations around the asset)

The rule-based algorithm stabilizes around one solution

,

where they are, or move keeping the symmetry around

the asset and s annin the whole set of solutions

We have never seen the symmetric motion in

simulations; in practice it can be ruled out (note: the

symmetric motion may be a plus, not a minus!)


37/39

Some more commentsSome more comments

Small comms overheadEach vehicle communicates with its closest neighbor

Data to be TX:

Agent Position Maximum Detection Sonar Range

Built in Emergency Procedure

If an a ent loses comms oes to the asset

Distributed, scalable algorithm, independent from AUV #

Comms delay do not alter result, but imply longer vehicle


38/39

Experimental testExperimental test

6-30 September 10, Pianosa Island

Networked communicationCooperative localization

Secur ty e av our

2 vehicles


39/39

ConclusionsConclusions

A set of tools for autonomous cooperative adaptive

context: data-driven adaptation

deterministic and probabilistic metrics

acoustic communication prediction

cooperative distributed localization

graph-theoretic approach

guaranteed optimality

ess ex e, commun ca on n ens ve

behaviour-basedrobust

g comms

optimality and convergence not guaranteed but for special cases

cooperative auv 2

Documents