From Swarm Intelligence to From Swarm Intelligence to Swarm EngineeringSwarm Engineering

Alan FT WinfieldAlan FT WinfieldIntelligent Autonomous Systems LabIntelligent Autonomous Systems Lab

www.www.iasias..uweuwe.ac..ac.ukuk

out of the lab and into the real world

This talkThis talk

Questions:Questions:How can we design swarm intelligence in a How can we design swarm intelligence in a methodologically rigorous way?methodologically rigorous way?How can we formally prove or validate swarm How can we formally prove or validate swarm engineered systems?engineered systems?

This talkThis talkThe IAS labThe IAS labCase study: a wireless connected swarmCase study: a wireless connected swarmSwarm EngineeringSwarm Engineering

The IAS LaboratoryThe IAS Laboratory

Swarm RoboticsSwarm Robotics

[Melhuish]

[Wessnitzer, Melhuish]

Collective sorting

Emergent formation

A Lighter Than Air Web ServerA Lighter Than Air Web Server

[Welsby]

The Flying FlockThe Flying Flock

[Welsby, Melhuish, Winfield]

Energy Autonomy: Energy Autonomy: SlugBotSlugBot

[Kelly, Holland ]

Energy Autonomy: Energy Autonomy: EcoBotEcoBot

[Greenman, Melhuish, Ieropoulos]

The The WhiskerbotWhiskerbot

www.www.whiskerbotwhiskerbot.org.org[Melhuish, Pipe, Pearson, Gilhespy et al]

A case study in Swarm A case study in Swarm Robotics: Robotics: hypothesishypothesis……

““That it is possible to maintain swarm integrity That it is possible to maintain swarm integrity using wireless networking aloneusing wireless networking alone””In other words:In other words:

Is it possible to use wireless networking as a Is it possible to use wireless networking as a structural componentstructural component in building multiin building multi--robot robot systems..?systems..?

We seek simple rules linking locomotion with We seek simple rules linking locomotion with communicationscommunications

To create emergent swarm coherence and To create emergent swarm coherence and Scalable control of swarm morphologyScalable control of swarm morphology

[Nembrini, PhD Thesis 2004]

A Minimalist ApproachA Minimalist Approach

Robots haveRobots haveRange limited, omniRange limited, omni--directional wireless directional wireless communicationscommunications

Situated communicationsSituated communicationsRobots can transmit their identityRobots can transmit their identityBut signal strength not availableBut signal strength not available

No global positional informationNo global positional informationNo range or bearing sensorsNo range or bearing sensorsOnly local knowledge of network topologyOnly local knowledge of network topology

Primitive behaviour: iPrimitive behaviour: i

1. Connected 2. Connection Lost

Continuous PING: send Are You There, respond with Yes I’m Here

Primitive behaviour: iiPrimitive behaviour: ii

3. Turn Back 4. Reconnected, choose New Heading

Continuous PING: send Are You There, respond with Yes I’m Here

Basic AlgorithmBasic Algorithm

Extend the basic primitive to multiple Extend the basic primitive to multiple robotsrobots……

React to the number of neighbours in range, React to the number of neighbours in range, i.e. the number of connections i.e. the number of connections KK

Swarm divisionSwarm division

But the basic algorithm cannot prevent But the basic algorithm cannot prevent swarm divisionswarm division……

Shared Neighbour AlgorithmShared Neighbour AlgorithmIf you lose a connection to robot If you lose a connection to robot NN, find how many of , find how many of your still connected neighbours have your still connected neighbours have NN in their neighbour in their neighbour lists: lists: nSharednShared

If If nSharednShared drops below drops below ββ, turn back, turn backIf If KK is rising, choose new random headingis rising, choose new random heading

Area ControlArea Control

The single parameter The single parameter ββ controls determines the controls determines the swarm coverageswarm coverage

Area control examplesArea control examples

Swarm disposition for Swarm disposition for ββ = 1, = 1, ββ = 4= 4

Directed SwarmingDirected Swarming

Consider now the problem of directing the Consider now the problem of directing the swarm (taxis) toward a beaconswarm (taxis) toward a beacon

We could introduce differential sensing into We could introduce differential sensing into one individualone individual

But this is highly dependent on signalBut this is highly dependent on signal--toto--noise rationoise ratio……and completely fails to exploit the spatial and completely fails to exploit the spatial distribution of the swarmdistribution of the swarm

Instead give each robot a simple binary Instead give each robot a simple binary sensor (illuminated or not illuminated)sensor (illuminated or not illuminated)

Emergent swarm taxisEmergent swarm taxis

For the illuminated For the illuminated (red) robots set the (red) robots set the value of value of ββ to infinityto infinity

The red robots then The red robots then shrink together to form shrink together to form a complete grapha complete graphReds become blues, Reds become blues, which become more which become more mobile, resulting inmobile, resulting in……slow translation toward slow translation toward the beaconthe beacon

beacon

Red: illuminated

Blue: occluded

Swarm taxis with obstaclesSwarm taxis with obstacles

Introduce occluding Introduce occluding obstaclesobstacles

The swarm finds it The swarm finds it way between the way between the narrow obstaclesnarrow obstacles

beacon

Encapsulation of the beaconEncapsulation of the beacon• An unexpected emergent phenomenon…

Swarm morphology controlSwarm morphology controlBy introducing a differential velocity By introducing a differential velocity between illuminated and occluded robots between illuminated and occluded robots we have emergent morphology controlwe have emergent morphology control

Emergent concentric symmetryEmergent concentric symmetry

2 cell types

3 cell types

EmergentEmergentradial radial symmetrysymmetry

Physical ImplementationPhysical ImplementationExperimental platform: the LinuxBotExperimental platform: the LinuxBot

Play: n2th2n7th2c50

What is a What is a Dependable SwarmDependable Swarm??It is a complex distributed system, designed using the It is a complex distributed system, designed using the Swarm IntelligenceSwarm Intelligence paradigm, which meets paradigm, which meets standardsstandardsof analysis, design and test that would give sufficient of analysis, design and test that would give sufficient confidence that the system could be employed in confidence that the system could be employed in critical applicationscritical applications

Q: What are these standards? Q: What are these standards? A: They don't existA: They don't exist

The purpose of our current work is to develop a The purpose of our current work is to develop a framework for the analysis, design and test of framework for the analysis, design and test of dependable swarmsdependable swarmsI propose to call this framework Swarm EngineeringI propose to call this framework Swarm Engineering

[Winfield et al, LNCS 3342, 2005]

Analysis Design Test

Assurance of DependabilityAssurance of Dependability

What makes swarm engineered systems What makes swarm engineered systems different?different?

System functionality achieved through emergenceSystem functionality achieved through emergenceSwarms are dynamical, stochastic, nonSwarms are dynamical, stochastic, non--linear linear systemssystemsTask completion Task completion becomes very hard to define.becomes very hard to define.

Designing the SwarmDesigning the Swarm

Structured Design Methodology

Use Waterfall (v-shaped) model?

Robot design

Swarm designProblematical because there are (as yet) noprincipled approaches to the design ofemergence

Ideally we need a formal, provableapproach to the design of individualswithin the swarm

Swarm design and robotdesign are tightly coupled

(Structured) Swarm Engineering(Structured) Swarm Engineering

Swarm Design

Robot Implementation

Swarm Test

Swarm Analysis

Swarm Test Specification

Requirements Specification Dependable Swarm

Top downFunctional Decomposition

Bottom upIntegration

and Test

Robot Design / Analysis Robot Test

Robot Design Specification

RTS

Single Agent Engineering

Morphology/Behaviours

Working Robots

Code

Simulation

(Dynamic) Data Flow Diagram(Dynamic) Data Flow Diagram

Robot 4

Data (Message) FlowsBetween neighbours

WirelessRange

Robot 3

Robot 5

Robot 2

Robot 1

Single Robot ProcessesSingle Robot Processes

Behaviour-basedControlProcess

UDPMessage

ServerNeighbourhoodConnectivity

Messages from Neighbours

Messages toNeighbours

Level 1 processLevel 2 process

We extend Lyapunov stability theory to We extend Lyapunov stability theory to secondsecond--order stability theoremsorder stability theorems

then use the partial subsumption relationship between then use the partial subsumption relationship between the 1the 1stst and 2and 2ndnd order Lyapunov stability theorems as the order Lyapunov stability theorems as the basis for a formal model of the subsumption basis for a formal model of the subsumption architecturearchitecture

Provably StableProvably StableBehaviourBehaviour--based Controlbased Control

Avoidance Behaviour

Network Behaviour

S

ActuatorsColony-style control architecture

W t 0 W t Wmax W t Wmax 0

W t 0

Direct Lyapunov DesignDirect Lyapunov DesignWe use the 2We use the 2ndnd order Lyapunov stability theorems as the order Lyapunov stability theorems as the basis for a design procedure for the motor schema of a basis for a design procedure for the motor schema of a behaviour modulebehaviour module

Model the Open-Loop Dynamics

Define goal state Sand its neighbourhood

For each point in the gridselect a control action

Define a piecewise mapfunction

and define a grid of pointsover the neighbourhood

select control actions that yield themost stabilising behaviour accordingto 2nd order stability theorems

in which grid points arethe central states of each i/o pair and their associatedselected actions are the function outputs

[Harper and Winfield, accepted for RAS]

Swarm modelling and analysisSwarm modelling and analysisLiveness

the property ofexhibiting desirable

behaviours

Safetythe property of not

exhibiting undesirablebehaviours

SimulationMathematical

Modelling

Single Robot

Multiple Robots

+

Hazard Analysis

Randomerrors

Systematic(design) errors

Single Robot

Multiple Robots

State Transition DiagramState Transition DiagramTurn back

Reverse

Random turn

All pathsblocked

Fwd blockedrear path clear

Obstacle left or right front

Swarm Lost

Swarm Found

Forward

Haltbac

Spin

Network BehaviourAvoidance Behaviour

ModellingModellingCurrent work is attempting to model the Current work is attempting to model the wireless connected swarm, by extending wireless connected swarm, by extending the probabilistic approach of the probabilistic approach of Martinoli Martinoli et et al.al.Take the Finite State MachineTake the Finite State Machine

then express as an ensemble of probabilistic then express as an ensemble of probabilistic FSMsFSMs::

Coherence Forward Avoid

The basic FSM

Probabilistic FSMProbabilistic FSM

Each box representsthe number of robotsin the swarm:• in a given state, and• with a given number of connectionsThe PFSM thus describes the state/ connection structure of the swarmUsing the modelling approach of Martinoli et al[IJR, 2004]

HazardsHazardsFailure Modes and Effects Analysis Failure Modes and Effects Analysis (FMEA)(FMEA)H1: Motor Failure, but all othersystems remain functional

The robot impedes the swarm fromtranslational motion (anchors theswarm)

H2: Avoid sensor failure The robot remains connected butavoid behaviour fails; the robot willoscillate between opposite edges ofthe swarm

H3: Communications failure The robot disconnects from the swarmand is eventually lost; it is treated bythe rest of the swarm as a dynamicobstacle

H4: All systems failed Robot is simply treated by rest ofswarm as a static obstacle

FSM with hazardsFSM with hazards

Coherence Forward Avoid

H1: motor failure

Pl Pa

PH1PH1 PH1

H2: Pa=0H3: Pl=1, Pr=0

Pr

H4: all systems failure

PH4

Using Temporal LogicUsing Temporal Logicto Specify Emergent Behavioursto Specify Emergent Behaviours

We are investigating the use of a Linear We are investigating the use of a Linear Time Temporal Logic to specify (and Time Temporal Logic to specify (and possibly prove) emergent propertiespossibly prove) emergent propertiesNASA have explored formal methods NASA have explored formal methods within the Autonomous within the Autonomous NanoNano--Technology Technology (ANTS) project ((ANTS) project (Rouff Rouff et al, 2004)et al, 2004)

however that work did not investigate a however that work did not investigate a temporal logictemporal logic

Swarm Swarm specificationspecification

Specify the safety andliveness properties of

each robot(in terms of lower level

behaviours)

Then specify the Swarmas the logical ‘and’ of all

the robots

Specification of Specification of Emergent Emergent PropertiesProperties

First specify the emergent properties

Now attempt to prove(or disprove) that the

swarm of robots satisfiesthe emergent behaviours

[Winfield, Sa et al, accepted for Taros 05]

Testing the SwarmTesting the Swarm

System Test(swarm)

Component Test(single robot)

Witness tests against aSystem Test Specification (STS)

Tests for Liveness

Tests for (partial) Safety

Tolerance and robustnessto random errors (and threats)

Dynamic/Static Analysis

+

Problematical because of theneed to create test harnesses

Testing the swarmTesting the swarmWe need toWe need to

establish robust measures for achievement of establish robust measures for achievement of desired (emergent) behaviours, thendesired (emergent) behaviours, thendefine (statistical) test for these measuresdefine (statistical) test for these measures

Qe – Mean quality of encapsulationRe – Mean radius of encapsulation

Vs – Mean swarmvelocity toward target

Frequency that Qe>Qthreshold in a given time period for given starting conditions

Swarm Tests in progressSwarm Tests in progress

Swarm tests can provide an Swarm tests can provide an environmentenvironment for single robot testfor single robot test

State;position;heading;sensor readings;connectivity

Actual behaviour

Expectedbehaviour

Pass/Fail

Controlled Swarm Tests

Swarm test results

Single robot simulation

Single Robot Tests

A roadmap towards swarm A roadmap towards swarm engineeringengineering

Substantial work is needed before Substantial work is needed before dependable swarms can become realitydependable swarms can become reality

We need to extend and strengthen analytical approaches to We need to extend and strengthen analytical approaches to modelling of swarm systemsmodelling of swarm systemsWe need to extend and strengthen formal approach to provably We need to extend and strengthen formal approach to provably stable intelligent controlstable intelligent control

To include safety as well as livenessTo include safety as well as livenessWe need a more principled approach to the design of emergenceWe need a more principled approach to the design of emergenceWe need to start work on 'safety' analysis at the swarm levelWe need to start work on 'safety' analysis at the swarm levelWe need to develop metrics, methodologies and practices for the We need to develop metrics, methodologies and practices for the testing of swarm engineered systemstesting of swarm engineered systems

DiscussionDiscussion

But... can or should we really think about But... can or should we really think about classical approaches to system validation classical approaches to system validation in the context of swarm engineering?in the context of swarm engineering?

some in classical safety systems believe the some in classical safety systems believe the standard approach is already breaking down standard approach is already breaking down for very complex (conventional) systemsfor very complex (conventional) systemsperhaps a new engineering paradigm calls for perhaps a new engineering paradigm calls for new approaches to dependability?new approaches to dependability?

IAS lab acknowledgementsIAS lab acknowledgementsProf Prof Owen HollandOwen HollandProfProf Andrew AdamatzkyAndrew AdamatzkyProf Chris MelhuishProf Chris MelhuishProf John Prof John GreenmanGreenmanDr Tony PipeDr Tony PipeDr Ben de Lacy CostelloDr Ben de Lacy CostelloDr Ian KellyDr Ian KellyDr Julien Dr Julien NembriniNembriniDr Jan Dr Jan WessnitzerWessnitzerDr Chris HarperDr Chris HarperIoannis IeropoulosIoannis IeropoulosJason Jason WelsbyWelsbyIan Ian HorsfieldHorsfieldIan Ian GilhespyGilhespy

And finally, back to the futureAnd finally, back to the future……

Bristol Pioneer, Dr W Grey WalterBristol Pioneer, Dr W Grey Walter

Machina Speculatrix