Objectives - Penn Engineeringesteager/Dynamic_Vehicle... · 2010-02-23 · 1 motion coordination:...
Transcript of Objectives - Penn Engineeringesteager/Dynamic_Vehicle... · 2010-02-23 · 1 motion coordination:...
Overview of MURI Activities:Coverage control, partitioning policies, pursuit
algorithms, and dynamic vehicle routing
Francesco Bullo
Center for Control,Dynamical Systems & Computation
University of California at Santa Barbara
http://motion.me.ucsb.edu
Review Meeting, ARO SWARMS MURI W911NF-05-1-0219University of Pennsylvania, February 23, 2010
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 1 / 53
Objectives
Develop broad set of coordination algorithms for autonomous networkedvehicles:
1 Coverage and surveillanceminimal gossip asynchronous comm
2 Pursuit strategies
3 Task allocation and dynamic vehicle routing
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 2 / 53
Technical Approach
Distributed algorithms
Graph theory and combinatorics
Computational geometry and geometric optimization
Nonlinear control theory
Queueing theory and stochastic processes
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 3 / 53
Major Breakthroughs
Coverage control
gossip communicationadaptive coverage and explorationboundary monitoring via synchronization
Cooperative pursuit strategies
pursuit with sensing limitations (limited range and range-only)pursuit with agility limitations
Task allocation and dynamic vehicle routingstatic and dynamic problemsgeometric heuristics and approximation algorithmsgrowing range of models and scenarios
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 4 / 53
DoD Benefits
Coverage and surveillance strategies
minimal communicationautomatic synchronization
Novel bio-inspired approaches to pursuit/evasion
strategies incorporating sensing/mobility constraints
Geometric algorithms for target assignment and vehicle routingprovably optimal or constant-factor under simplified assumptionsgrowing range of models and scenarios
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 5 / 53
SWARMS Collaborations
1 David Skelly: bio-inspired cooperative pursuit strategies
2 Daniela Rus and David Skelly: adaptive coverage and explorationStephen Smith (from UCSB PhD to postdoc with CSAIL @
MIT)
3 Ali Jadbabaie: discrete coverage control and network flow
4 Brian Anderson and Stephen Morse: network localization
5 Dan Koditschek: topology of the space of partitions
6 Emilio Frazzoli: dynamic vehicle routing
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 6 / 53
Success Stories
Plenary speaker: NESC 05, BMSC 06, HSCC 06, Int Conf Applied Math andComputing 08, IEEE MSC ’09, IFAC NECSYS ’10
Promoted to Full Professor & ME vicechair at UCSB
Elevated IEEE Fellow, 2010
Article selection for inclusion in SIGEST section of SIAM Review ’09
Outstanding Paper Award, IEEE Control Systems Magazine, 2008
Students recognition: Best student paper award at ACC06; finalist atCDC05, CDC07, and ACC’10. UCSB CCDC Best PhD Thesis. DoDSMART fellowship
6 PhD thesis on cooperative control: Sara Susca, Ketan Savla, AnuragGanguli, Giuseppe Notarstefano, Stephen Smith, Shaunak Bopardikar
Special Issue in SIAM Journal of Control and Optimization on Control andOptimization in Cooperative Networks, Jan 2009
Workshops: at Centro De Giorgi in Dec 2007, CDC in Dec 08, minitutorialat 2009 SIAM Conf on Dynamical Systems, BMSC 2009, ACC 2010
Open source software library for visibility computation, http://visilibity.org
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 7 / 53
Publications: Journal articles
1 S. Susca, S. Martınez, and F. Bullo. Monitoring environmental boundaries with a robotic sensor network. IEEETransactions on Control Systems Technology, 16(2):288–296, 2008
2 C. Gao, J. Cortes, and F. Bullo. Notes on averaging over acyclic digraphs and discrete coverage control. Automatica,44(8):2120–2127, 2008
3 K. Plarre and F. Bullo. On Kalman filtering for detectable systems with intermittent observations. IEEE Transactions onAutomatic Control, 54(2):386–390, 2009
4 S. L. Smith and F. Bullo. Monotonic target assignment for robotic networks. IEEE Transactions on Automatic Control,54(9):2042–2057, 2009
5 S. D. Bopardikar, F. Bullo, and J. P. Hespanha. On discrete-time pursuit-evasion games with sensing limitations. IEEETransactions on Robotics, 24(6):1429–1439, 2008
6 R. Carli and F. Bullo. Quantized coordination algorithms for rendezvous and deployment. SIAM Journal on Control andOptimization, 48(3):1251–1274, 2009
7 S. L. Smith and F. Bullo. The dynamic team forming problem: Throughput and delay for unbiased policies. Systems &Control Letters, 58(10-11):709–715, 2009
8 S. D. Bopardikar, F. Bullo, and J. P. Hespanha. A cooperative Homicidal Chauffeur game. Automatica, 45(7):1771–1777,2009
9 S. D. Bopardikar, S. L. Smith, F. Bullo, and J. P. Hespanha. Dynamic vehicle routing for translating demands: Stabilityanalysis and receding-horizon policies. IEEE Transactions on Automatic Control, January 2010. (Submitted, Mar 2009) toappear
10 M. Pavone, E. Frazzoli, and F. Bullo. Distributed and adaptive algorithms for vehicle routing in a stochastic and dynamicenvironment. IEEE Transactions on Automatic Control, August 2009. (Submitted, Apr 2009) to appear
11 G. Notarstefano and F. Bullo. Distributed abstract optimization via constraints consensus: Theory and applications. IEEETransactions on Automatic Control, October 2009. Submitted
12 F. Bullo, R. Carli, and P. Frasca. Gossip coverage control for robotic networks: Dynamical systems on the the space ofpartitions. SIAM Review, January 2010. Submitted
plus 25 conference articles and book chapters
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 8 / 53
Publications: Book on Robotic Networks
1 intro to distributed algorithms (graphtheory, synchronous networks, andaveraging algos)
2 geometric models and geometricoptimization problems
3 model for robotic, relative sensingnetworks, and complexity
4 algorithms for rendezvous,deployment, boundary estimation
Manuscript by F. Bullo, J. Cortes, and
S. Martınez. Princeton Univ Press, 2009,
ISBN 978-0-691-14195-4. Freely downloadable
at http://coordinationbook.info with
tutorial slides and software libraries.
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 9 / 53
Link to other efforts
ARO Institute for Collaborative Biotechnology, “Bio-inspiredStochastic Search and Decision Making for Robotic Networks”
NSF “Distributed Illumination Problems for Visually-guided Agents”
AFOSR MURI on “Decision Dynamics in Mixed Networks”
NSF “Minimalist Mapping and Monitoring”
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 10 / 53
Technology Transition
Established connections (conversations, mutual visits, planning):1 UtopiaCompression, Dr Jacob Yadegar2 Northrop Grumman, Dr Han Park3 PARC, Dr Haitham Hindi4 Mayachitra, Dr Marco Zuliani5 Los Alamost National Lab, Dr Misha Chertkov
PhD students in DoD labs1 Current PhD candidate Karl Obermeyer, SMART fellow at
UCSB/AFRL Wright Patterson
Graduate students in industry:1 Anurag Ganguli, UtopiaCompression2 Sara Susca, Honeywell Research Labs3 Nathan Owen, Boeing Space & Intelligence Systems
PhD students in academia:1 Ketan Savla, Research Scientist at MIT2 Stephen Smith, Postdoc at MIT with Daniela Rus3 Shaunak Bopardikar, Postdoc at UCSB4 Giuseppe Notarstefano, Professor at University of Lecce
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 11 / 53
Plans for Years 4-5
1 Research planned in proposal:
Design of motion coordination algorithms, e.g., deployment,connectivity maintenance, target assignmentCoordination with minimal (gossip, quantized, intermittent)communicationBioinspired pursuit behaviors
2 New unanticipated directions:
Distributed linear programming with applications to formation controland target localizationSynchronization for surveillanceQueueing theory for robotic networks
Francesco Bullo (UCSB) SWARMS MURI Overview 23feb10 @ Penn 12 / 53
Dynamic Vehicle Routing for Robotic Networks:Models, Fundamental Limits and Algorithms
Francesco Bullo
Center for Control,Dynamical Systems & Computation
University of California at Santa Barbara
http://motion.me.ucsb.edu
Review Meeting, ARO SWARMS MURI W911NF-05-1-0219University of Pennsylvania, February 23, 2010
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 13 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 14 / 53
Summary: Dynamic Vehicle Routing for Robotic Networks
Problem Setup:DVR = planning policies for vehicles tovisit targets arriving in real time
1 vehicle: dynamics, comm models
2 coordination: partitions/teaming
3 targets: priority, mobile
Technical Approach
models: dimensional analysis,intrinsic regimes, phase transitions
fundamental limits onperformance (delay vs throughput)
algorithm design: optimal orconstant-factor, adaptive,distributed
Publications and Dissemination
3 PhD Theses @ UCSB: K. Savla,S. L. Smith and S. D. Bopardikar
9 journal articles (6 in TAC, 1SICON, 1 SCL, 1 AIAA JGCD)
plenary at ’09 IEEE Conf ControlApps, tutorial at ’09 SIAM DynamSystems Conf, ACC’10 workshop
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 15 / 53
Acknowledgements
Ketan Savla(PhD @ UCSB, nowresearch scientist @
MIT)
Stephen Smith(PhD @ UCSB, now
postdoc @ MIT)
Marco Pavone(graduating PhD @
MIT)
Emilio Frazzoli (MIT)
Shaunak D. Bopardikar (UCSB) and Joao P. Hespanha (UCSB):translating targets
Ruggero Carli (UCSB), Joey W. Durham (UCSB), andPaolo Frasca (Universita di Roma): gossip coordination
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 16 / 53
From coordination and static routing toDynamic Vehicle Routing
Simple coordination problems arise in static environments
1 motion coordination: rendezvous, deployment, flocking
2 task allocation, target assignment3 static vehicle routing (P. Toth and D. Vigo ’01)
Routing policies vs planning algorithms
dynamic, stochastic and adversarial events take place
1 design policies (in contrast to pre-planned routes or motion planningalgorithms) to specify how to react to events
2 dynamic demands add queueing phenomena to the combinatorialnature of vehicle routing
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 17 / 53
Literature review
Shortest path through randomly-generated and worst-case points(Beardwood, Halton and Hammersly, 1959 — Steele, 1990)
Traveling salesman problem solvers (Lin, Kernighan, 1973)
DVR formulation on a graph (Psaraftis, 1988)
DVR on Euclidean plane (Bertsimas and Van Ryzin, 1990–1993)
Unified receding-horizon policy (Papastavrou, 1996)
Recent developments in DVR for robotic networks:
Adaptation and decentralization (Pavone, Frazzoli, FB: TAC, in press)
Nonholonomic / Dubins UAVs (Savla, Frazzoli, FB: TAC 2008)
Pickup delivery tasks (Waisanen, Shah, and Dahleh: TAC 2008)
Heterogeneous vehicles and team forming (Smith and Bullo: SCL 2009)
Distinct-priority targets (Smith, Pavone, FB, Frazzoli: SICON, in press)
Moving targets (Bopardikar, Smith, Hespanha, FB: TAC, in press)
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 18 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 19 / 53
Prototypical Dynamic Vehicle Routing Problem
Given:
a group of vehicles, and
a set of service demands
Objective:provide service in minimum timeservice = take a picture at location
Vehicle routing (All info known ahead of time, Dantzig ’59)
Determine a set of paths that allow vehicles to service the demands
Dynamic vehicle routing (New info in real time, Psaraftis ’88)
New demands arise in real-time
Existing demands evolve over time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 20 / 53
Prototypical Dynamic Vehicle Routing Problem
Given:
a group of vehicles, and
a set of service demands
Objective:provide service in minimum timeservice = take a picture at location
Vehicle routing (All info known ahead of time, Dantzig ’59)
Determine a set of paths that allow vehicles to service the demands
Dynamic vehicle routing (New info in real time, Psaraftis ’88)
New demands arise in real-time
Existing demands evolve over time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 20 / 53
Prototypical Dynamic Vehicle Routing Problem
Given:
a group of vehicles, and
a set of service demands
Objective:provide service in minimum timeservice = take a picture at location
Vehicle routing (All info known ahead of time, Dantzig ’59)
Determine a set of paths that allow vehicles to service the demands
Dynamic vehicle routing (New info in real time, Psaraftis ’88)
New demands arise in real-time
Existing demands evolve over time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 20 / 53
Prototypical Dynamic Vehicle Routing Problem
Given:
a group of vehicles, and
a set of service demands
Objective:provide service in minimum timeservice = take a picture at location
Vehicle routing (All info known ahead of time, Dantzig ’59)
Determine a set of paths that allow vehicles to service the demands
Dynamic vehicle routing (New info in real time, Psaraftis ’88)
New demands arise in real-time
Existing demands evolve over time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 20 / 53
Prototypical Dynamic Vehicle Routing Problem
Given:
a group of vehicles, and
a set of service demands
Objective:provide service in minimum timeservice = take a picture at location
Vehicle routing (All info known ahead of time, Dantzig ’59)
Determine a set of paths that allow vehicles to service the demands
Dynamic vehicle routing (New info in real time, Psaraftis ’88)
New demands arise in real-time
Existing demands evolve over time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 20 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 21 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Plain-vanilla re-optimization?
Example: DVR on segment
Objective: minimize averagewaiting time
Strategy: re-optimize at eachevent
10 0.5
1 For adversarial target generation, vehicle travels forever without everservicing any request =⇒ unstable queue of outstanding requests
2 Even if queue remains bounded, what about performance? how farfrom the optimal?
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 22 / 53
Online algorithms?
Online algorithms (Jaillet and M. R. Wagner ’06)
online algorithm operates based on input information up to thecurrent time
online algorithm is r -competitive if
Costonline ≤ rCostoptimal offline(I ), ∀ problem instances I .
Disadvantages
1 cumulative cost
2 worst-case analysis
3 not possible to include a-priori information (e.g., arrival rate)
4 not as clear what competitive ratio means
5 so far, only few simple DVR problems admit online algorithms
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 23 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 24 / 53
Algorithm design
1 Distinct light load and high load behaviors
2 For light load, optimal assets placement for fastest response3 For high load,
1 solve combinatorial optimization over available info2 repeat in receding-horizon fashion
4 For multi-agent problems,1 (often) optimal load balancing via territory partitioning2 (other times) teaming
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 25 / 53
Algorithm design
1 Distinct light load and high load behaviors
2 For light load, optimal assets placement for fastest response3 For high load,
1 solve combinatorial optimization over available info2 repeat in receding-horizon fashion
4 For multi-agent problems,1 (often) optimal load balancing via territory partitioning2 (other times) teaming
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 25 / 53
Algorithm design
1 Distinct light load and high load behaviors
2 For light load, optimal assets placement for fastest response3 For high load,
1 solve combinatorial optimization over available info2 repeat in receding-horizon fashion
4 For multi-agent problems,1 (often) optimal load balancing via territory partitioning2 (other times) teaming
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 25 / 53
Algo #1: Receding-Horizon Shortest-Path policy
Receding-Horizon Shortest-Path (RH-SP)
For η ∈ (0, 1], single agent performs:
1: while no customers, move to center2: while customers waiting
1 compute shortest path through current targets
2 service η-fraction of path
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 26 / 53
Algo #1: Receding-Horizon Shortest-Path policy
Receding-Horizon Shortest-Path (RH-SP)
For η ∈ (0, 1], single agent performs:
1: while no customers, move to center2: while customers waiting
1 compute shortest path through current targets
2 service η-fraction of path
M. Pavone, E. Frazzoli, and F. Bullo. Distributed and adaptive algorithms for vehiclerouting in a stochastic and dynamic environment. IEEE Transactions on AutomaticControl, August 2009. (Submitted, Apr 2009) to appear
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 26 / 53
RH-SP analysis
Implementation:
NP-hard computation, but effective heuristics
Stability:1 queue is stable if service time < interarrival time
2 service time =length shortest path(n)
n(n = # customers)
3 queue is stable if (length of shortest path(n)) = sublinear f(n)
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 27 / 53
RH-SP analysis
Implementation:
NP-hard computation, but effective heuristics
Stability:1 queue is stable if service time < interarrival time
2 service time =length shortest path(n)
n(n = # customers)
3 queue is stable if (length of shortest path(n)) = sublinear f(n)
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 27 / 53
RH-SP analysis
Implementation:
NP-hard computation, but effective heuristics
Stability:1 queue is stable if service time < interarrival time
2 service time =length shortest path(n)
n(n = # customers)
3 queue is stable if (length of shortest path(n)) = sublinear f(n)
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 27 / 53
RH-SP analysis
Implementation:
NP-hard computation, but effective heuristics
Stability:1 queue is stable if service time < interarrival time
2 service time =length shortest path(n)
n(n = # customers)
3 queue is stable if (length of shortest path(n)) = sublinear f(n)
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 27 / 53
RH-SP analysis
Implementation:
NP-hard computation, but effective heuristics
Stability:
1 queue is stable if service time < interarrival time
2 service time =length shortest path(n)
n(n = # customers)
3 queue is stable if (length of shortest path(n)) = sublinear f(n)
Combinatorics in Euclidean space (Steel ’90)
Worst-case and expected bounds
length shortest path(n) ≤ βworst
√n
limn→+∞
length shortest path(n) = βexpected
√n
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 27 / 53
RH-SP analysis: continued
Adaptation: the policy does not require knowledge of
1 vehicle velocity v , environment Q
2 arrival rate λ and spatial density function f
3 expected on-site service s
Performance:
1 in light load, delay is optimal
2 in high load, delay is within a multiplicative factor from optimal
3 multiplicative factor depends upon f and is conjectured to equal 2
no known adaptive algo with better performancevery little known outside of asymptotic regimes
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 28 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 29 / 53
Euclidean TSP and Dubins TSP
Euclidean TSP (ETSP)
NP-hard
effective heuristics available
length(ETSP) ∈ O(√
n)
Dubins TSP (DTSP)Given a set of points find the shortest tour with bounded curvature
not a finite dimensional problem
no prior algorithms or results (asof 2006)
length(DTSP) sub-linear in n ?
K. Savla, E. Frazzoli, and F. Bullo. Traveling Salesperson Problems for the Dubinsvehicle. IEEE Transactions on Automatic Control, 53(6):1378–1391, 2008
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 30 / 53
Stochastic DTSP
Problem Statement Given a set of n independently and uniformlydistributed points, design polynomial-time algorithm with smallestexpected DTSP tour length
Theorem: For n iid uniformly distributed points:
E[length of DTSP(n)] ∼ n2/3
Lower bound proof based on “area of reachable set”
1 area of reachable set in time t by Dubins with radius ρ is t3
3ρ
2 expected distance to nearest target (n iid uniform targets) is 34(3ρ
n )1/3
3 lenght of tour cannot be less than n times this distance
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 31 / 53
Constructive upper boundbased on environment tiling tuned to vehicle dynamics
10
!
p! p+B!(!)
!
Fig. 2. Construction of the “bead” B!(!). The figure shows how the upper half of the boundary is constructed, the bottom half is symmetric.
Next, we study the probability of targets belonging to a given bead. Consider a bead B entirely contained in Q
and assume n points are uniformly randomly generated in Q. The probability that the ith point is sampled in B is
µ(!) =Area(B!(!))
Area(Q).
Furthermore, the probability that exactly k out of the n points are sampled in B has a binomial distribution, i.e.,
indicating with nB the total number of points sampled in B,
Pr[nB = k| n samples] =!
n
k
"µk(1! µ)n!k.
If the bead length ! is chosen as a function of n in such a way that " = n · µ(!(n)) is a constant, then the limit
for large n of the binomial distribution is [31] the Poisson distribution of mean ", that is,
limn"+#
Pr[nB = k| n samples] ="k
k!e!" .
C. The Recursive Bead-Tiling Algorithm
In this section, we design a novel algorithm that computes a Dubins path through a point set in Q. The proposed
algorithm consists of a sequence of phases; during each of these phases, a Dubins tour (i.e., a closed path with
bounded curvature) will be constructed that “sweeps” the set Q. We begin by considering a tiling of the plane such
June 30, 2006 DRAFT
Q
Key properties of the bead
1 Beads tile the plane
2 Approaching and leaving a bead horizontally, Dubins can service a target
first analysis of joint combinatorics, dynamics and stochasticextensions to STLC systems by Itani, Dahleh and Frazzoli
extensions to multi-vehicle Dubins
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 32 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 33 / 53
Algo #2: Load balancing via territory partitioning
RH-SP + Partitioning
Each agent i:
1: computes own cell vi in optimal partition2: applies RH-SP policy on vi
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 34 / 53
Territory partitioning akin to animal territory dynamics
Tilapia mossambica, “Hexagonal
Territories,” Barlow et al, ’74
Red harvester ants, “Optimization, Conflict, and
Nonoverlapping Foraging Ranges,” Adler et al, ’03
Sage sparrows, “Territory dynamics in a sage sparrows
population,” Petersen et al ’87
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 35 / 53
Optimal partitioning cost functions
Expected wait time (light load problem)
H(p, v) =
∫v1
‖q − p1‖dq + · · ·+∫vn
‖q − pn‖dq
n robots at p = {p1, . . . , pn}environment is partitioned into v = {v1, . . . , vn}
H(p, v) =n∑
i=1
∫vi
f (‖q − pi‖)φ(q)dq
φ : R2 → R≥0 density
f : R≥0 → R penalty function
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 36 / 53
From optimality conditions to algorithms
H(p, v) =n∑
i=1
∫vi
f (‖q − pi‖)φ(q)dq
Theorem (Alternating Algorithm, Lloyd ’57)
1 at fixed positions, optimal partition is Voronoi
2 at fixed partition, optimal positions are “generalized centers”
3 alternate v-p optimization=⇒ local optimum = center Voronoi partition
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 37 / 53
Gossip partitioning policy
1 Random communication between two regions2 Compute two centers3 Compute bisector of centers4 Partition two regions by bisector
F. Bullo, R. Carli, and P. Frasca. Gossip coverage control for robotic networks: Dynam-ical systems on the the space of partitions. SIAM Review, January 2010. Submitted
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 38 / 53
Gossip partitioning policy: sample implementation
Player/Stage platform
realistic robot models in discretized environments
integrated wireless network model & obstacle-avoidance planner
J. W. Durham, R. Carli, P. Frasca, and F. Bullo. Discrete partitioning and cover-age control with gossip communication. In ASME Dynamic Systems and ControlConference, Hollywood, CA, October 2009
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 39 / 53
Gossip partitioning policy: analysis results
1 class of dynamical systems on space of partitionsi.e., study evolution of the regions rather of the agents
2 convergence to centroidal Voronoi partitions (under mild conditions)
3 novel results in topology, analysis and geometry:1 compactness of space of finitely-convex partitions with respect to the
symmetric difference metric2 continuity of various geometric maps (Voronoi as function of
generators, centroid location as function of set, multicenter functions)3 LaSalle convergence theorems for dynamical systems on metric
spaces with deterministic and stochastic switches
conjectures about topology of space of partitionsasymmetric gossip algorithms, akin to stigmergy
tolerance to failures, arrivals, and dynamic environments
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 40 / 53
Asymmetric gossip partitioning policy
Asymmetric communication akin to animal stigmergy
1 unidirectional delayed randomized links
2 use covering instead of partitions
3 doubly-greedy strategy:improve performance & diminish overlap
Asymmetric gossip partitioning algorithmRandom directed communication from region vi to region vj :
1 Compute two centers
2 Compute bisector of centers3 Perform two operations on region j
1 add every point in region vi that is closer to jth center2 del every point in region vj ∪ vi that is closer to ith center
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 41 / 53
Today’s Outline
1 IntroductionSummaryFrom coordination and static routing to Dynamic Vehicle RoutingLiterature review
2 Prototypical Dynamic Vehicle Routing ProblemBasic modelPreliminary attemptsAlgorithm design and analysis
3 DVR Extensions, Variations, Applications, . . .Vehicle dynamics: Dubins TSP problemVehicle coordination: Gossip partitioningTargets: Moving targetsTargets: Heterogeneous targets requiring teamsTargets: Priority levels
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 42 / 53
DVR for moving targets
Very little is know about moving targets:
1 no polynomial time algorithms for shortest path
2 no length estimates
3 no efficient DVR algorithms
S. D. Bopardikar, S. L. Smith, F. Bullo, and J. P. Hespanha. Dynamic vehicle routing fortranslating demands: Stability analysis and receding-horizon policies. IEEE Transactions onAutomatic Control, January 2010. (Submitted, Mar 2009) to appear
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 43 / 53
Translating targets
Problem parameters:
speed ratio v :
v =target speed
vehicle speed
arrival rate λ
segment width W
deadline distance L
W
L
L = +∞ L is finiteStabilize queue Maximize capture fraction
v < 1 translational path policy translational path policy
v ≥ 1 Not possible for any λ > 0 longest path policy
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 44 / 53
Translating targets
Problem parameters:
speed ratio v :
v =target speed
vehicle speed
arrival rate λ
segment width W
deadline distance L
W
L
L = +∞ L is finiteStabilize queue Maximize capture fraction
v < 1 translational path policy translational path policy
v ≥ 1 Not possible for any λ > 0 longest path policy
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 44 / 53
Shortest Translational Path Fraction (v < 1)
Shortest translational path fraction policy
1 Compute shortest translational path through all demands higherthan L/2.
2 Service demands on path for min of: L/(2v) time units, and time totravel entire path
3 Repeat
Translational path computation(Hammar and Nilsson, 2002)
Order: scaled shortest static pathMotion: intercept on straight line
L/2
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 45 / 53
Maximizing capture fraction for v ≥ 1
For v ≥ 1, it is optimal to remain on deadline
Reachable targets
Reachability graph is directed and acyclic
Longest path can be computed in polynomial time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 46 / 53
Maximizing capture fraction for v ≥ 1
For v ≥ 1, it is optimal to remain on deadline
Reachability graph is directed and acyclic
Longest path can be computed in polynomial time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 46 / 53
Maximizing capture fraction for v ≥ 1
For v ≥ 1, it is optimal to remain on deadline
Reachability graph is directed and acyclic
Longest path can be computed in polynomial time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 46 / 53
Maximizing capture fraction for v ≥ 1
For v ≥ 1, it is optimal to remain on deadline
Reachability graph is directed and acyclic
Longest path can be computed in polynomial time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 46 / 53
Maximizing capture fraction for v ≥ 1
For v ≥ 1, it is optimal to remain on deadline
Reachability graph is directed and acyclic
Longest path can be computed in polynomial time
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 46 / 53
Longest path policy (v ≥ 1)
Longest path (LP) policy
1 Compute reachability graph ofunserviced demands
2 Compute longest path in graph
3 Capture first demand on path byintercepting on deadline
4 Repeat
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 47 / 53
Summary of boundary guarding
L = +∞ L is finiteStabilize queue Maximize capture fraction
v < 1
v ≥ 1 Not possible for any λ > 0
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 48 / 53
DVR for moving targets
Results (limits and constant-factor analysis removed for brevity):
problem structure: four regimes
fundamental limits on stability/capture fraction
efficient algorithms in all four regimes
L = +∞ L is finiteStabilize queue Maximize capture fraction
v < 1 Translational path policy Translational path policy
v ≥ 1 Not possible for any λ > 0 Longest path policy
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 49 / 53
Studying more general scenarios
Relaxed assumptions:
Non-Poisson
Non-uniform
Different speeds
Different directions
Finite capture radius
More general setup:
Higher dimensions
Advance information
very little known about generic target motion models
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 50 / 53
Targets: Heterogeneous targets requiring teams
Problem setup:
Heterogeneous vehicles
Tasks require vehicle teams
Goal: Minimize task delay
Consider only unbiased policies:Equal expected delay to all tasks
Provably efficient policies in certain scenariosVery rich problem
S. L. Smith and F. Bullo. The dynamic team forming problem: Throughput anddelay for unbiased policies. Systems & Control Letters, 58(10-11):709–715, 2009
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 51 / 53
Targets: Priority levels
Problem setup:
n vehicles
Two classes of tasks α, β
α – high priorityβ – low priority
Goal: minimize cDα + (1− c)Dβ
c ∈ (0, 1) gives bias toward α
c = 0.80p = 0.82
Provably efficient policy
Extends to m classes
S. L. Smith, M. Pavone, F. Bullo, and E. Frazzoli. Dynamic vehicle routing withpriority classes of stochastic demands. SIAM Journal on Control and Optimization,48(5):3224–3245, 2010
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 52 / 53
Today: Dynamic Vehicle Routing for Robotic Networks
Problem Setup:DVR = planning policies for vehicles tovisit targets arriving in real time
1 vehicle: dynamics, comm models
2 coordination: partitions/teaming
3 targets: priority, mobile
Technical Approach
models: dimensional analysis,intrinsic regimes, phase transitions
fundamental limits onperformance (delay vs throughput)
algorithm design: optimal orconstant-factor, adaptive,distributed
Publications and Dissemination
3 PhD Theses @ UCSB: K. Savla,S. L. Smith and S. D. Bopardikar
9 journal articles (6 in TAC, 1SICON, 1 SCL, 1 AIAA JGCD)
plenary at ’09 IEEE Conf ControlApps, tutorial at ’09 SIAM DynamSystems Conf, ACC’10 workshop
Francesco Bullo (UCSB) Dynamic Vehicle Routing 23feb10 @ Penn 53 / 53