From Swarm Intelligence to Swarm Engineering

1
From Swarm Intelligence to Swarm Engineering
out of the lab and into the real world

• Alan FT Winfield
• Intelligent Autonomous Systems Lab
• www.ias.uwe.ac.uk

2
This talk

• Questions
• How can we design swarm intelligence in a
  methodologically rigorous way?
• How can we formally prove or validate swarm
  engineered systems?
• This talk
• The IAS lab
• Case study a wireless connected swarm
• Swarm Engineering

3
The IAS Laboratory
4
Swarm Robotics
Collective sorting
Melhuish
Emergent formation
Wessnitzer, Melhuish
5
A Lighter Than Air Web Server
Welsby
6
The Flying Flock
Welsby, Melhuish, Winfield
7
Energy Autonomy SlugBot
Kelly, Holland
8
Energy Autonomy EcoBot
Greenman, Melhuish, Ieropoulos
9
The Whiskerbot

• www.whiskerbot.org

Melhuish, Pipe, Pearson, Gilhespy et al
10
A case study in Swarm Robotics hypothesis

• That it is possible to maintain swarm integrity
  using wireless networking alone
• In other words
• Is it possible to use wireless networking as a
  structural component in building multi-robot
  systems..?
• We seek simple rules linking locomotion with
  communications
• To create emergent swarm coherence and
• Scalable control of swarm morphology

Nembrini, PhD Thesis 2004
11
A Minimalist Approach

• Robots have
• Range limited, omni-directional wireless
  communications
• Situated communications
• Robots can transmit their identity
• But signal strength not available
• No global positional information
• No range or bearing sensors
• Only local knowledge of network topology

12
Primitive behaviour i
1. Connected
2. Connection Lost
Continuous PING send Are You There, respond with
Yes Im Here
13
Primitive behaviour ii
3. Turn Back
4. Reconnected, choose New Heading
Continuous PING send Are You There, respond with
Yes Im Here
14
Basic Algorithm

• Extend the basic primitive to multiple robots
• React to the number of neighbours in range, i.e.
  the number of connections K

15
Swarm division

• But the basic algorithm cannot prevent swarm
  division

16
Shared Neighbour Algorithm

• If you lose a connection to robot N, find how
  many of your still connected neighbours have N in
  their neighbour lists nShared
• If nShared drops below ß, turn back
• If K is rising, choose new random heading

17
Area Control

• The single parameter ß controls determines the
  swarm coverage

18
Area control examples

• Swarm disposition for ß 1, ß 4

19
Directed Swarming

• Consider now the problem of directing the swarm
  (taxis) toward a beacon
• We could introduce differential sensing into one
  individual
• But this is highly dependent on signal-to-noise
  ratio
• and completely fails to exploit the spatial
  distribution of the swarm
• Instead give each robot a simple binary sensor
  (illuminated or not illuminated)

20
Emergent swarm taxis

• For the illuminated (red) robots set the value of
  ß to infinity
• The red robots then shrink together to form a
  complete graph
• Reds become blues, which become more mobile,
  resulting in
• slow translation toward the beacon

beacon
Red illuminated Blue occluded
21
Swarm taxis with obstacles
beacon

• Introduce occluding obstacles
• The swarm finds it way between the narrow
  obstacles

22
Encapsulation of the beacon

• An unexpected emergent phenomenon

23
Swarm morphology control

• By introducing a differential velocity between
  illuminated and occluded robots we have emergent
  morphology control

24
Emergent concentric symmetry
2 cell types
3 cell types
25
Emergentradial symmetry
26
Physical Implementation

• Experimental platform the LinuxBot

Play n2th2 n7th2c50
27
What is a Dependable Swarm?

• It is a complex distributed system, designed
  using the Swarm Intelligence paradigm, which
  meets standards of analysis, design and test that
  would give sufficient confidence that the system
  could be employed in critical applications
• Q What are these standards?
• A They don't exist
• The purpose of our current work is to develop a
  framework for the analysis, design and test of
  dependable swarms
• I propose to call this framework Swarm Engineering

Winfield et al, LNCS 3342, 2005
28
Assurance of Dependability
Analysis
Design
Test

• What makes swarm engineered systems different?
• System functionality achieved through emergence
• Swarms are dynamical, stochastic, non-linear
  systems
• Task completion becomes very hard to define.

29
Designing the Swarm
Structured Design Methodology
Use Waterfall (v-shaped) model?
Problematical because there are (as yet)
no principled approaches to the design
of emergence
Swarm design
Robot design
Ideally we need a formal, provable approach to
the design of individuals within the swarm
Swarm design and robot design are tightly coupled
30
(Structured) Swarm Engineering
Requirements Specification
Dependable Swarm
Simulation
Swarm Test Specification
Swarm Test
Swarm Design
Swarm Analysis
Robot Design Specification
Working Robots
Robot Design / Analysis
Robot Test
RTS
Bottom up Integration and Test
Top down Functional Decomposition
Morphology/Behaviours
Code
Robot Implementation
Single Agent Engineering
31
(Dynamic) Data Flow Diagram
Robot 5
Robot 1
Robot 3
Robot 2
Data (Message) Flows Between neighbours
Robot 4
Wireless Range
32
Single Robot Processes
Messages to Neighbours
Behaviour- based Control Process
Messages from Neighbours
UDP Message Server
Neighbourhood Connectivity
Level 1 process Level 2 process
33
Provably StableBehaviour-based Control

• We extend Lyapunov stability theory to
  second-order stability theorems
• then use the partial subsumption relationship
  between the 1st and 2nd order Lyapunov stability
  theorems as the basis for a formal model of the
  subsumption architecture

Network Behaviour
Avoidance Behaviour
S
Colony-style control architecture
Actuators
34
Direct Lyapunov Design

• We use the 2nd order Lyapunov stability theorems
  as the basis for a design procedure for the motor
  schema of a behaviour module

Model the Open- Loop Dynamics
Define goal state S and its neighbourhood
and define a grid of points over the neighbourhood
For each point in the grid select a control action
select control actions that yield the most
stabilising behaviour according to 2nd order
stability theorems
in which grid points are the central states of
each i/o pair and their associated selected
actions are the function outputs
Define a piecewise map function
Harper and Winfield, accepted for RAS
35
Swarm modelling and analysis
Liveness the property of exhibiting
desirable behaviours
Safety the property of not exhibiting
undesirable behaviours

Hazard Analysis
Mathematical Modelling
Simulation
Random errors
Systematic (design) errors
Single Robot
Multiple Robots
Single Robot
Multiple Robots
36
State Transition Diagram
Turn back
Swarm Lost
Swarm Found
Random turn
Network Behaviour Avoidance Behaviour
All paths blocked
Fwd blocked rear path clear
Obstacle left or right front
Reverse
b
a
c
Spin
37
Modelling

• Current work is attempting to model the wireless
  connected swarm, by extending the probabilistic
  approach of Martinoli et al.
• Take the Finite State Machine
• then express as an ensemble of probabilistic FSMs

Coherence
Forward
Avoid
The basic FSM
38
Probabilistic FSM

• Each box represents
• the number of robots
• in the swarm
• in a given state, and
• with a given number of connections
• The PFSM thus describes the state/ connection
  structure of the swarm
• Using the modelling approach of Martinoli et al
• IJR, 2004

39
Hazards

• Failure Modes and Effects Analysis (FMEA)

40
FSM with hazards
H2 Pa0
H3 Pl1, Pr0
Pl
Pa
Coherence
Forward
Avoid
Pr
PH1
PH1
PH1
H1 motor failure
PH4
H4 all systems failure
41
Using Temporal Logicto Specify Emergent
Behaviours

• We are investigating the use of a Linear Time
  Temporal Logic to specify (and possibly prove)
  emergent properties
• NASA have explored formal methods within the
  Autonomous Nano-Technology (ANTS) project (Rouff
  et al, 2004)
• however that work did not investigate a temporal
  logic

42
Swarm specification
Specify the safety and liveness properties
of each robot (in terms of lower
level behaviours) Then specify the Swarm as the
logical and of all the robots
43
Specification of Emergent Properties
First specify the emergent properties Now
attempt to prove (or disprove) that the swarm of
robots satisfies the emergent behaviours
Winfield, Sa et al, accepted for Taros 05
44
Testing the Swarm
System Test (swarm)
Component Test (single robot)

Witness tests against a System Test Specification
(STS)
Dynamic/Static Analysis
Problematical because of the need to create test
harnesses
Tests for Liveness
Tests for (partial) Safety
Tolerance and robustness to random errors (and
threats)
45
Testing the swarm

• We need to
• establish robust measures for achievement of
  desired (emergent) behaviours, then
• define (statistical) test for these measures

Vs Mean swarm velocity toward target
Qe Mean quality of encapsulation Re Mean
radius of encapsulation
Frequency that QegtQthreshold in a given time
period for given starting conditions
46
Swarm Tests in progress
47
Swarm tests can provide an environment for single
robot test
Controlled Swarm Tests
Single Robot Tests
Statepositionheading sensor readingsconnectivi
ty
Single robot simulation
Expected behaviour
Actual behaviour
Pass/Fail
Swarm test results
48
A roadmap towards swarm engineering

• Substantial work is needed before dependable
  swarms can become reality
• We need to extend and strengthen analytical
  approaches to modelling of swarm systems
• We need to extend and strengthen formal approach
  to provably stable intelligent control
• To include safety as well as liveness
• We need a more principled approach to the design
  of emergence
• We need to start work on 'safety' analysis at the
  swarm level
• We need to develop metrics, methodologies and
  practices for the testing of swarm engineered
  systems

49
Discussion

• But... can or should we really think about
  classical approaches to system validation in the
  context of swarm engineering?
• some in classical safety systems believe the
  standard approach is already breaking down for
  very complex (conventional) systems
• perhaps a new engineering paradigm calls for new
  approaches to dependability?

50
IAS lab acknowledgements

• Prof Owen Holland
• Prof Andrew Adamatzky
• Prof Chris Melhuish
• Prof John Greenman
• Dr Tony Pipe
• Dr Ben de Lacy Costello
• Dr Ian Kelly
• Dr Julien Nembrini
• Dr Jan Wessnitzer
• Dr Chris Harper
• Ioannis Ieropoulos
• Jason Welsby
• Ian Horsfield
• Ian Gilhespy

51
And finally, back to the future

• Bristol Pioneer, Dr W Grey Walter

Machina Speculatrix