Presentation @ WOA 2015

Self-Organising UAVs for Wide Area Fault-tolerant Aerial Monitoring

Self-Organising UAVs for Wide AreaFault-tolerant Aerial Monitoring

Massimiliano De Benedetti, Fabio D’Urso, Fabrizio Messina,Giuseppe Pappalardo, Corrado Santoro

ARSLAB - Autonomous and Robotic Systems LaboratoryDipartimento di Matematica e Informatica

Universita di Catania, Italy

WOA 2015 - Napoli, Italy, Sept. 17-18, 2015De Benedetti, D’Urso, Messina, Pappalardo, Santoro. Self-Organising UAVs - WOA 2015 1


Outline

1 Aim and Scope

2 System Model

3 Area Coverage AlgorithmsOverlay ConstructionFlock FormationPath Planning

4 Simulation Study

5 Conclusions

De Benedetti, D’Urso, Messina, Pappalardo, Santoro. Self-Organising UAVs - WOA 2015 2


Aim and Scope

Aim and Scope

Context: Aerial inspection of a wide area of terrain formonitoring purposes.

Autonomy: ability to perform the required inspection withoutexternal control.

Requirements: to minimise mission time and to ensure ahigh degree of fault tolerance

PROPOSAL: To use of set of Unmanned Aerial Vehicles (VTOLmultirotors) able to:

self-organise in a flock

cover the desired area

handle faults of one or more UAVsDe Benedetti, D’Urso, Messina, Pappalardo, Santoro. Self-Organising UAVs - WOA 2015 3


System Model

System Model

The system is made of:

A set of mobile entities (agents), the multirotors

A certain physical area to be covered

An inertial 3-dimensional reference system: latitude,longitude, altitude, heading

The area to be monitored can be represented with a shape of arectangle: (x1, y1), (x2, y2), (x3, y3), (x4, y4).



System Model

System Model—Agents

Each UAV/agent is equipped with:

A flight stabilization systemA GPS-based autopilotA low-power wireless transmission system (with limitedrange)A monitoring sensor (i.e. a camera) able to cover, at analtitude of z , the area (wsz , hsz)

At run-time, each UAV/agent is characterised by a unique ID anda time-dependent pose:

posei (t) = (xi , yi , zi , headi )



System Model

System Model—Mission

A mission is characterised by:

The area to be covered, i.e. : (x1, y1), (x2, y2), (x3, y3), (x4, y4)

The number of agents, N

The operating altitude, Z

The objective is:

to ensure the complete coverage of the area

to minimise mission time

to avoid over-coverage



Area Coverage Algorithms

Area Coverage

To perform area coverage, three algorithms are employed:

Overlay Construction, agents self-organise in an overlaynetwork in order to know each other and be able to interact

Flock Formation, agents self-organise in order to assume aflock shape with certain characteristics

Flock Driving, agent flock sets-up and follows a certain pathin order to cover the mission area




Overlay Construction

Overlay Construction—(I)

Each agent k maintains a local Agent Database (ADB) whichstores the information of the other agents known by k :

ADB = {(IDi , posei ,HOP(k)i ,TSi )}

where:

IDi is the identifier of agent i

posei is the pose of agent i

HOP(k)i is the number of hops—in the view of k—required to

reach i

TSi is the timestamp (in the localtime of i) at which poseihas been generated





Overlay Construction—(II)

Periodically, agent k retrieves its position posek and send inbroadcast the information:

ADB ∪ {(IDk , posek , 0,TSk)}

Since transmission system is limited in range, only some of theother agents will receive the message.

A so-formed message (we call it RADB) received by an agent k∗ isthen analised by k∗ and compared with its local ADB.





Overlay Construction—(III)

On the reception of RADB, the following operation are performedby agent k∗:

each hop count field in RADB is incremented

for each tuple ∈ RADB, find ID in ADB of k∗

if it is not found, ADB ← ADB ∪ {tuple}if it is found then

if the TS (RADB) > TS (ADB), ADB ← ADB ∪ {tuple}else keep information in ADB and discard tuple

After some iterations each agent will know:

all the other agents of the system;

their last position (which is periodically updated);

its “distance” with respect to all the other agents.




Flock Formation

Flock Formation—(I)

Aim: to maximize area coverage and mimizeover-covering.

We choose a flock shape which ensures that a certain are isnot monitored by two agents.

The ideal shape is a linear placement along a formation line,perpendicular to the direction of flight.

The distance between two agents is kept close to wsZ (withan overlap of 20%—configurable).




Flock Formation

Flock Formation—(II)

To form a flock, a leader is elected as the agent with thelowest ID.

Each agent elects the leader by directly choosing it from theADB, no further interaction is required.

As soon as the overlay construction algorith proceeds, allagents will elect the same leader.

Tuples in the ADB are subject to aging: if a tuple is notupdated within a timeout, it is discarded.

Should the leader fail, its data (in the ADBs) will bediscarded and a new leader will be elected.

The leader establishes the flock formation line which is thevirtual line perpendicular to leader heading.




Flock Formation

Flock Formation—(III)

Rule R1: Separation

The geographic distance di to each other neighbour agent i iscomputed.

If di is less than a hard threshold DH, no motion is applied(the UAV remains in hovering for the current iteration); this isrequired to avoid any possible collision.

If di is less than a soft threshold DS > DH, a repulsion forceis generated for the agent, with an heading always parallel tothe formation line.




Flock Formation

Flock Formation—(IV)

Rule R2: Alignment

The distance di to each other neighbour agent i is computed.

The agent k ′ which is the nearest to the leader (w.r.t. hi ) isfound (or the leader itself).

The orientation of k ′ is copied by using a yaw rotation.




Flock Formation

Flock Formation—(V)

Rule R3: Cohesion

The distance di to each other neighbour agent i is computed.

The agent k ′ which is the nearest to the leader (w.r.t. hi ) isfound (or the leader itself).

Agent k ′ is approached by means of translated flight (if theyare too close, rule R1 intervenes)




Path Planning

Path Planning—(I)

To cover the desired area, the following algorithm is employed:

1 The overall area is subdivided in small portions.

2 Periodically, all the agents spread the data about the portionsalready covered by each of them.

3 The leader, on the basis of such data, computes the optimalpath and follows it driving the entire flock.




Path Planning

Path Planning—(II)

The area is subdivided before starting the mission as follows:A roto-translation of the reference is performed in order to set onecorner of the rectangle area to the origin.The flight is performed by keeping the formation line parallel tothe X axis.The area is discretized in stripes along Y axis using a size of hsZper stripe. Each stripe is numbered.




Path Planning

Path Planning—(III)

Each agent holds and updates an Area Parts Database APDB whichcontains the information about the parts already covered. It is a set of:

(StripeID, {(XS1,XE1), (XS2,XE2), . . . , (XSn,XEn)})

Each stripe part is represented by its starting and ending X coordinate(XSi ,XEi ).Two close parts are always merged together by means of a stripe unionoperation:




Path Planning

Path Planning—(IV)

The leader periodically starts a distributed aggregatedquery which aims at obtaining the knowledge of the APDB ofall the other agents

Agent respond by broadcasting their APDB

An agent receiving an other’s APDB, merges received datawith its own APDB and broadcasts the resulting set

...




Path Planning

Path Planning—(V)

...

The leader, on the basis of received data, computes somepossible paths and chooses the one currently nearest to theflock

The path is then smoothed

The leader starts to follow the chosen path and drives theoverall flock.



Simulation Study

Simulation Study

To validate the proposed approach, a software simulator has beenbuilt:

written ad-hoc in C++ with QT libraries

able to simulate all the algorithms by graphically showing theflight of the UAVs and colouring the parts of the area covered

able to simulate faults in one or more agents (crash) or in thecommunication system (packet loss)

The simulator is available on request.



Conclusions

Conclusions and Future Work

While preliminary tests showed the validity of the approach,further simulation studies are needed in order to numericallyevaluate:

fault tolerance degreemission time trend vs. number of agentspercentage of portions over-covered due to faults....

This work is part of a research project (PON-CLARA) for themonitoring and control of landslips.

A further step will be the implementation on real UAVs



Conclusions

Self-Organising UAVs for Wide AreaFault-tolerant Aerial Monitoring

Massimiliano De Benedetti, Fabio D’Urso, Fabrizio Messina,Giuseppe Pappalardo, Corrado Santoro

ARSLAB - Autonomous and Robotic Systems LaboratoryDipartimento di Matematica e Informatica

Universita di Catania, Italy

WOA 2015 - Napoli, Italy, Sept. 17-18, 2015De Benedetti, D’Urso, Messina, Pappalardo, Santoro. Self-Organising UAVs - WOA 2015 23

Presentation @ WOA 2015

Education

Transcript of Presentation @ WOA 2015