Cooperative control of multi-vehicle

systems

Sonia Martnez

Mechanical and Aerospace Engineering

Jacobs School of Engineering
University of California at San Diego
http://faemino.ucsd.edu/soniamartinez/index.html

B-DRASS seminar, The Boeing Company

May 20, 2009, Huntington Beach, CA

Collaborators: Andrew Kwok, Minghui Zhu

Francesco Bullo, Jorge Cortes, Emilio Frazzoli
Multi-agent networks

What kind of systems?

Groups of agents with control, sensing, communication and computing

Each individual
senses its immediate environment
communicates with others
processes information gathered
takes local action in response

,
Self-organized behaviors in biological groups

Natural multi-agent systems:

,
Decision making in animals

Able to
deploy over a given region
assume specified pattern
rendezvous at a common point
jointly initiate motion/change direction
in a synchronized way

Species achieve synchronized behavior

with limited sensing/communication between individuals

without apparently following group leader

(Couzin et al, Nature 05; Conradt et al, Nature 03)

,
 Songhwai Oh, Phoebus Chen, Bruno its potential importance to the North American climate.
4 http://webs.cs.berkeley.edu
Ubiquitous TanyaInstrumentation
We have recently shown that more than 75% of the here have been very few campaigns to examine how

Sinopoli, Roosta, Sastry

black carbon over the west coast of North America was particles transported across the Pacific Ocean influence
the result of transport across the Pacific from East Asia clouds and radiative forcing in the region. In April 2004, the

Emerging technologies and multi-agent systems

Network
UnderstandingEmbedded
phenomena: Systems (NEST)
and other regions ((ADLEYETAL 2006) during springtime
and possibly other seasons as well. We have also
shown the dust-soot mixture has a large impact on TOA
and surface radiative forcing and atmospheric solar
Cloud Indirect Effects Experiment (CIFEX) (2AMANATHAN
2003; HTTPBORNETTOUCSDEDUCIFEX+ING!IR?REQUEST?3)/5#3$
PDF) followed the Intercontinental Transport2006
Transformation (ITCT) campaign, undertook a pilot
and American
Chemical Control Conference, Minneapolis

Group heating rates. However, we do not know the magnitude examination with airborne aerosol and cloud instruments
Data collection for offline analysis of the impact on clouds through the indirect effect, a onboard the University of Wyomings +ING!IR aircraft. The
Environmental monitoring, habitat monitoring [Szewczyk
potentially important mechanism for regulating cloud collected data provided unique insights into the role of long-
range transport of aerosols across the Pacific Ocean (2OBERTS
forcing. 4HE POTENTIAL IMPACT OF THESE SOOT DUST MODIlED
http://webs.cs.berkeley.edu
et al., 2004] FORCINGTERMSONTHEREGIONALCIRCULATIONANDCLIMATE is one
major motivating factor for this proposed study.
Structural monitoring [Pakzad et al., 2005]
ETAL 2006; (ADLEYETAL 2006; 7ILCOXETAL 2006). First,
(ADLEY ET AL 2006 demonstrated that aerosols from long-
range transport were the dominant source of black carbon
and other fine particles above 2 km over the west coast of

Distributed habitat/environment monitoring systems

2006 American Control Conference, Minneapolis
North America during spring time (Figure 6). Second, the
size distribution of these particles had markedly different
characteristics when compared with marine sources or North
American sources (Figure 7). Lastly, the particles provided
efficient source of CCN (2OBERTSETAL 2006) and nucleated

25 Motes on
Ubiquitous Instrumentation
cloud drops that in turn reduced the drop radii (Figure 8) and
enhanced the cloudy sky albedo (7ILCOXETAL 2006).

Damaged sidewall
Understanding phenomena:
Data collection for offline analysis
Soil monitoring Soil monitoring
Environmental monitoring, habitat monitoring [Szewczyk
Sensor Webs Everywhere et al., 2004]
Ubiquitous Instrumentation
Understanding phenomena:
Structural monitoring [Pakzad et al., 2005]

Great DuckforIsland
Data collection offline analysis
Understanding phenomena:
Environmental monitoring, habitat monitoring
Data collection for [Szewczyk et al., 2004] Scripps multi-UAV
The UAV fleet at )BOJNBBEIPP Airport, the Maldives duringproject,
the MAC UCSD ASAP project illustration
offline analysis Redwoods
campaign in March 2006. Wind Response Figure 6: CFORS fraction of Asian BC to total BC as a function
Structural monitoring [Pakzad et al., 2005] of altitude at 130W ((ADLEYETAL., 2006).
Environmental monitoring, habitat monitoring [SzewczykOf Golden70!#0AGE Gate Bridge
Vineyards
et al., Disaster relief operations, rapid deployment systems, surveillance
Detecting
2004] changes in the environment: 2006 American Control Conference, Minneapolis
Thresholds, phase transitions, anomaly detection
25 Motes on
Structural monitoring [Pakzad et al., 2005] Damaged sidewall
Security systems, surveillance [Brooks et al., 2004;
Health monitoring
Arora et al., 2004], health care in smart bridges, buildings, roads
Wildfire detection [Doolin, Sitar, 2005]
Fault detection, threat detection
Intel Research Soil monitoring
Health Care 11
25 Motes on
Damaged sidewall
Great Duck Island

Redwoods Wind Response

Fire Soil
Response
monitoring Soil monitoring
Of Golden Gate Bridge
2006 American Control Conference, Minneapolis
Vine
surveillance Berkeley motes
2006 American Control Conference, Minneapolis
Sandia National Labs Berkeley

Great Duck Island

,
Research challenges

What useful engineering tasks can be performed

with limited-sensing/communication agents?

Dynamics simple interactions give rise to

Feedback
Information flow
Reliability/performance
rich emerging behavior
rather than open-loop computation
for known/static setup
who knows what, when, why, how,
dynamically changing
robust, efficient, predictable behavior

How to coordinate individual agents into coherent whole?

Objective: systematic methodologies to design and analyze

cooperative strategies to control multi-agent systems

,
 2EGIONAL)MPACTONTHE0ACIlC #)&%84HE,ONG2ANGE
AND.ORTH!MERICAN#LIMATE 4RANSPORTOF$USTAND!EROSOLS
Research program: What are we after?
/ur focus is on the Pacific Ocean, largely because of
ACROSSTHE0ACIlC/CEAN
its potential importance to the North American climate.
We have recently shown that more than 75% of the 4here have been very few campaigns to examine how
black carbon over the west coast of North America was particles transported across the Pacific Ocean influence
the result of transport across the Pacific from East Asia clouds and radiative forcing in the region. In April 2004, the
and other regions ((ADLEYETAL 2006) during springtime Cloud Indirect Effects Experiment (CIFEX) (2AMANATHAN
and possibly other seasons as well. We have also 2003; HTTPBORNETTOUCSDEDUCIFEX+ING!IR?REQUEST?3)/5#3$
shown the dust-soot mixture has a large impact on TOA PDF) followed the Intercontinental Transport and Chemical
Development of system-theoretic tools to analyze general
and surface radiative forcing and atmospheric solar
heating rates. However, we do not know the magnitude
Transformation (ITCT) campaign, undertook a pilot
examination with airborne aerosol and cloud instruments
of the impact on clouds through the indirect effect, a onboard the University of Wyomings +ING!IR aircraft. The
coordination algorithms under disturbances
potentially important mechanism for regulating cloud
forcing. 4HE POTENTIAL IMPACT OF THESE SOOT
DUST MODIlED
collected data provided unique insights into the role of long-
range transport of aerosols across the Pacific Ocean (2OBERTS
FORCINGTERMSONTHEREGIONALCIRCULATIONANDCLIMATE is one ETAL 2006; (ADLEYETAL 2006; 7ILCOXETAL 2006). First,
major motivating factor for this proposed study.
Synthesis of distributed motion coordination algorithms
that are flexible, reconfigurable, scalable and verifiably correct
(ADLEY ET AL 2006 demonstrated that aerosols from long-
range transport were the dominant source of black carbon
and other fine particles above 2 km over the west coast of
North America during spring time (Figure 6). Second, the
size distribution of these particles had markedly different
characteristics when compared with marine sources or North
American sources (Figure 7). Lastly, the particles provided
efficient source of CCN (2OBERTSETAL 2006) and nucleated
cloud drops that in turn reduced the drop radii (Figure 8) and
enhanced the cloudy sky albedo (7ILCOXETAL 2006).

The UAV fleet at )BOJNBBEIPP Airport, the Maldives during the MAC
campaign in March 2006. Figure 6: CFORS fraction of Asian BC to total BC as a function
of altitude at 130W ((ADLEYETAL., 2006).

70!#0AGE

Synthesis of distributed filters and learning algorithms

to allow for distributed and adaptive decision making
Testbed development for technology transition
,
Text: Distributed Control of Robotic Networks

1 intro to distributed algorithms

(graph theory, synchronous
networks, and averaging algos)
2 geometric models and geometric
optimization problems
3 model for robotic, relative sensing
networks, and complexity
4 algorithms for rendezvous,
deployment, boundary estimation

Status: Freely downloadable at

http://coordinationbook.info
with tutorial slides & software libraries.
Shortly on sale by Princeton Univ Press

,
Outline

1 Motivation

2 Sample motion coordination algorithms and results

Multi-vehicle Deployment & Task Assignment
Multi-vehicle Rendezvous
Motion coordination for optimal estimation

3 Conclusions and Future Work

,
Deployment algorithms

Goal: Given n agents in an environment Q,

achieve optimal coverage by designing pi = ui , i {1, . . . , n}

Metrics capturing good deployment

Area jointly covered limited sensing ranges
Expected network cost to visit locations in Q:
constrained dynamics (motion constraints, environmental flows)
limited energy budget for each agent to move

Local Rules can be synthesized that are:

1 locally efficient, distributed, lead to stable emerging behaviors

2 robust to failures and asynchronous computations

,
Rendezvous algorithms

Goal: Given nodes (p1 , . . . , pn ) moving in environment Q,

gather at a given location of Q

Metrics capturing rendezvous objective

Disagreement in position
Diameter of network

Local Rules can be synthesized that:

1 Achieve task under switching connectivity conditions

2 Totally distributed
3 Work under rather persistent communication/sensing failures
4 Are robust to measurement noise and asynchronous behavior

,
Motion coordination for optimal estimation

Goal 1: Track a moving target q0

Goal 2: Track an evolving boundary

Information metrics
Fisher Information metrics
Goodness-of-interpolation metrics

Local Rules can be synthesized that are:

1 Distributed, able to track a global/local maxima (bounded error
depending on the velocity of change)
2 Robust to communication/sensing failures
3 Robust to asynchronous behaviors

,
Sample simulations

,
Summary

Basic motion coordination algorithms for multiagent systems

sample primitives: deployment, rendezvous, optimal estimation

System-theoretic tools for design:

proximity graphs, graph theory
distributed algorithms
set-valued maps, invariance principles
...

Features of algorithms:
scalable, adaptive, verifiably correct, robust, distributed

,
Current and Future work

Energy-efficient coordination algorithms

Motion coordination partially known, time-varying environments

Distributed optimization and learning for multi-agent systems

Selfish and cooperative

multi-agent behaviors
Distributed
receding-horizon control
policies

Coordination of multiple vehicles under communication limitations

Adversarial multi-robot teams

,
Thank you!