0% found this document useful (0 votes)
65 views48 pages

Motion Planning

This document discusses motion planning methods for multi-robot systems. It begins by noting the challenges in providing a comprehensive overview and focuses on foundational centralized and decentralized approaches. For centralized methods, it distinguishes between coupled approaches that search the combined state space and decoupled approaches that plan for each robot individually and then coordinate. Decentralized methods are classified based on the information available to each robot. The document then surveys examples of both centralized and decentralized planning techniques.

Uploaded by

Jasón De Elea
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
65 views48 pages

Motion Planning

This document discusses motion planning methods for multi-robot systems. It begins by noting the challenges in providing a comprehensive overview and focuses on foundational centralized and decentralized approaches. For centralized methods, it distinguishes between coupled approaches that search the combined state space and decoupled approaches that plan for each robot individually and then coordinate. Decentralized methods are classified based on the information available to each robot. The document then surveys examples of both centralized and decentralized planning techniques.

Uploaded by

Jasón De Elea
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 48

Motion Planning in

Multi-Robot Systems

Kostas E. Bekris
Department of Computer Science
Rutgers University
07/16/2015
Note to Online Reader
•  It is difficult to provide a comprehensive coverage of all motion
planning methods for multi-robot systems

•  An effort was made to cover foundational work in the case of


centralized solutions

•  For decentralized solutions, the presentation highlights methods that


the author has utilized in his research

•  But the version of the presentation on the TC’s website can


potentially be a live document that gets updated given your feedback
– So, if you believe that a certain line of work should be highlighted
here please contact Kostas Bekris (kostas.bekris @ cs.rutgers.edu)
Proposed Classification

Key question:
•  What information does an approach access?
-  Global: Centralized approaches
-  Local: Decentralized approaches
Important In & Beyond Robotics

Mul$ple  Direct  Applica$ons  


(including  centralized  methods)  
 
•  Warehouse  management  
•  Transporta$on  applica$ons  
Kiva  Systems/Amazon  
•  Controlling  teams  of  robots  
“Cossacks:  Back  to  War”  Game  
in  structured  environments  
•  Digital  entertainment  
•  Product  assembly  
•  Combinatorial  puzzles  and  
pure  scien$fic  curiosity  
Centralized Approaches

Key questions for centralized approaches:


•  What is the space that the method searches over for a solution?
-  Composite state space of all robots: Coupled approaches
-  Individual robot conf. space and coordination: Decoupled approaches
•  What kind of guarantees can be provided?
-  Safety, Completeness, Optimality
Decentralized Approaches

Key questions for decentralized approaches:


•  How does a local method access information about other robots?
-  Sensing or communication
-  Inference or shared information
•  What kind of properties can be provided?
-  Collision Avoidance, Deadlock/Livelock Avoidance
Centralized – Coupled Planning
Key features of Coupled Approaches

•  Consider the composite state space

•  Search can be performed with standard single-robot motion


planning methods in X, e.g.,
– combinatorial planners in low-dimensional cases,
– sampling-based planners, [Svestka and Overmars, 1998]
– optimization methods,
– search (A*) etc.

•  Then, it is possible to achieve the same properties as the algorithm


achieves in the single-robot case

But… computational issues!


Complexity Results

•  A complete algorithm [Schwartz and Sharir, 1983]


–  Coordinating planar disk-robots: Exponential complexity in the number of
robots

•  Exponential running time in some cases is unavoidable


–  Rectangular robots in rectangular region: PSPACE-hard [Hopkroft, Schwartz and Sharir, 1984]
–  NP-hard for disc robots in a simple polygon workspace [Spirakis,Yap 1984]
–  For 2-3 robots, reduce number of DOFs but computing paths where the robots
maintain contact [Aronov et al. 1999]
Unlabeled Case
•  A variation of the problem with
interchangeable robots [Kloder and Hutchinson 2005]
–  Group of identical robots that need to reach a set
of target positions

•  Could it be that it is an easier challenge?


–  No, unit-square robots moving amidst polygonal
obstacles and other variations are PSPACE-hard
[Solovey, Halperin RSS 2015]

•  Study of the disc robot case among polygonal obstacles:


–  Efficient solution when aiming for minimizing the longest robot path length
[Turpin, Michael and Kumar 2013]
•  The space must be star-shaped surrounding each start and target position
–  This has been relaxed to less restrictive sparsity requirement [Adler et al. 2014]
–  Efficient algorithm also in the case of minimizing total path length [Solovey et al. RSS
2015]
Centralized – Decoupled Planning
Basic Idea
•  First compute individual path for each robot
–  i.e., in the corresponding configuration space Ci

•  Then consider plan interactions to produce a solution that is


(hopefully) valid in the composite space X Decoupled    
Planning  
Proper=es  
•  When successful…
–  They solve problems orders of magnitude
faster than coupled alternatives!

•  But when the pair-wise interactions


are considered, the available choices
are already constrained…
– i.e., no completeness or optimality
guarantees in the general case
Prioritized Planning
•  Compute paths sequentially for different agents in order of priority
–  Higher-priority agents are considered moving obstacles for lower-priority one
[Erdmann and Lozano-Perez, 1986]

4  
2   1   •  Choice of priorities has significant
3   impact on solution quality [van den
Berg and Overmars, 2005]

•  Searching the space of priorities


can improve performance [Bennewitz,
Burgard,Thrun 2002]

Incremental  methods:    
•  plan  path  for  a  robot,  considering  the  paths  of  a  subset  of  the  other  agents  
•  a  plan-­‐merging  scheme  coordinates  ac$ons  to  detect  deadlocks  
•  when  a  circular  dependency  is  detected,  a  couple  planner  is  invoked  
[Alami  et  al.  1995,  Qutub  et  al.  1997]    
Velocity Tuning
Two step approach:
–  Fix paths for all agents and then in order of priority apply velocity tuning
•  i.e., select velocity for low priority agent along path so as to avoid collisions
•  treat high-priority agents as dynamic obstacles [Kant, Zucker 1986]

Idea relates to coordination


diagrams which were
developed for dual-arm
manipulation: [O’Donnell, Lozano-Perez 1989]
[Simeon, Leroy, Laumond 2002]

[Planning  Algorithms  -­‐  LaValle  2005]  

Extended to systems with more complex dynamics [Peng and Akell 2005]
Example Use of Velocity Tuning
Fixed Roadmaps
•  More flexible solutions if the robots are not constrained on
individual paths but on entire roadmaps [Ghrist, O’Kane and LaValle 2005]
–  Give rise to interesting coordination spaces (cube complexes)
–  Makes more sense to aim for Pareto optimal solutions

•  Similar idea:
–  Try to compute multiple diverse paths first for each agent [Green, Kelly 2007] [Knepper,
Mason 2009][Voss, Moll, Kavraki 2015]
–  Or make sure you are covering many different homotopic classes [Bhattacharya,
Kumar, Likhachev 2010]
Centralized
Discrete Case and New Insights
Difficult even in Discrete Domains
Remove the complexity of reasoning about the geometry
–  Employ a graph-based abstraction

The  problem  is  studied  in  many  different  


communi$es  under  different  names:  
•  Mul$-­‐agent  Planning  
•  Coopera$ve  Path  Finding  
•  Pebble  Mo$on  on  a  Graph  
•  Mul$-­‐agent  Naviga$on  

Finding optimal solutions is an NP-complete problem [Ratner and Warmuth, 1986]


Fast but Incomplete Methods
•  Computa$onally  efficient.    
•  Decoupled  framework.  
•  No  guarantees  for   [Silver  2005]  
–  Completeness.  
–  Path  Quality.  

[Sturtevant  and  Buro  2006]  


•  Dynamic  priori$za$on  and  windowed  search    
[Silver  2005]  

•  Spa$al  abstrac$on  with  heuris$c  computa$on    


[Sturtevant  and  Buro  2006]   [Wang  and  Botea  2008]  

•  Use  of  a  flow  network  with  replanning    


[Wang  and  Botea  2008]  

•  Smart  direc$on  maps  that  learns  movements    


[Jansen  and  Sturtevant  2008]  

[Jansen  and  Sturtevant  2008]  


Suboptimal but Complete Methods
•  S$ll  efficient:  polynomial  running  $me.  
•  They  will  find  a  solu$on  if  one  exists.  
•  They  do  not  provide  op$mal  paths.   [Peasgood,  Clark  et  al.  2008]  

•  Specific topologies
[Peasgood et al. 2008][Surynek 2009] [Wang  and  Botea  2011]  

•  Slideable grid-based problems 


[Wang and Botea 2011]
[Surynek  2009]  

•  Complete on trees 


[Khorshid et al. 2011]

•  “Push and Swap”: Polynomial-time solution  [Khorshid  et  al.  2011]  
on graphs with two empty vertices 
[Luna and Bekris 2011]

“Push  and  Swap”  SoIware  Package  Available:  


 Scales  up  to  Thousands  of  Agents  
[Luna  and  Bekris  2011]  
Foundations in Algorithmic Theory
•  Polynomial time feasibility test algorithm for graphs graphs [Kornhauser et al.
1984][Roger and Helmert 2012]

•  Linear time feasibility algorithm on trees [Auletta et al. 1999]

•  Linear algorithm for graphs with two blanks [Goraly and Hassin 2010]
Interesting Disparity

Finding   [Kron$ris,  Luna,  


Evalua$ng   Subop$mal   Bekris  SoCS  ‘13]  
Finding  an  
Feasibility   Op$mal  Path  
Paths   [Yu,  ‘13]  
Linear     Time!   Cubic     Time  
Extension  to  simultaneous  mo$on   NP-­‐hard  
[Yu,  Rus,  WAFR  ‘14]  
New Optimal Methods
•  Provide  path  quality  guarantees.  
[Wagner  and    
•  Coupled  framework  -­‐    oaen  A*-­‐based.   Choset  2013]  

•  Great  recent  progress  but…  scalability    


condi$onal  to  the  hardness  of  the  problem  

•  Op$mal  decoupling  
         [van  den  Berg  et  al.  RSS  2009]  
 

•  Working  on  independent  subproblems  


[Standley  2010,  Standley  and  Korf  2011]  
 

•  Subdimensional  expansion  search  space  


[Wagner  and  Choset  2011,  2013]  
[Standley  2010,    
Standley  and  Korf  2011]  
•  Conflict-­‐based  Search  
[Sharon,  Stern,  Sturtevant  2012,  2015]  

•  Cast  challenge  to  another  NP-­‐hard  problem  


–  Linear  Programming  [Yu,  LaValle  2013]  
–  Or  other  formal  methods  [Erdem  et  al.  2013,  Surynek  2012]  
[Yu  and  LaValle  2013]  
Back to Continuous Problems
•  Integrating sampling-based algorithms with pebble graph solvers to
address continuous challenges [Solovey and Halperin WAFR 12]

Discrete  RRT:  Integrated  the  ideas  


of  M*  with  RRT  for  solving  
con$nuous  problems  [Solovey,  
Salzman  and  Halperin  2014]  

•  We have recently transferred the idea in the context of rearranging


multiple movable bodies with a manipulator 
[Krontiris, Shome, Dobson, Kimmel Bekris Humanoids 2014] [Krontiris, Bekris RSS 2015]
[KronYris,  Shome,  Dobson,  Kimmel  Bekris  Humanoids  2014]    
[KronYris,  Bekris  RSS  2015]  
Multi-Arm Manipulation

[Koga,  Latombe  1994]  


Multi- [Cohen,  Philips,  Likhachev  RSS  2014  ]  
Arm
Planning [Dobson,  Bekris  IROS  2015]  

Planning  handoffs  and  stable  grasps  


Decentralized Approaches
Deconfliction for First-order Systems
It  is  possible  to  employ  
reac$ve  collision  avoidance  
methods  
•  No  need  to  employ  
communica$on  
 
e.g.  Reciprocal  Velocity  
Obstacles        
[van  den  Berg,  Lin,  Manocha  ‘08]  
   

a          b  
vb  
b  
VOa|b  
Deconfliction for First-order Systems

Reciprocal  Velocity  Obstacles        


[van  den  Berg,  Lin,  Manocha  ‘08]  
 
Extended  to  address  team  
coherence  constraints  
[Kimmel,  Bekris  AAMAS  ‘12]  
   

a          b  
vb  
b  
VOa|b  
Deadlock Issues
•  A prototypical motion coordination challenge
– Agent A must decide whether to move down Corridor 1 or 2.
– Similarly, Agent B will need to decide the same.

Corridor  2  
– G  
B A   GA  

Corridor  1   B  

– Assume employment of RVOs for safety purposes


– How can we achieve progress?
•  No communication, only observe the other agents
Motion Coordination Challenge

a2  
C(a2)  

GB   A   GA  

a1  
C(a1)  

For each agent, the cost of each action α is defined as C(α), the
length of the corresponding path to the goal.
Interaction Costs

a2  

GB   A   GA  

a1  
B   b2  
I11   b1  

Let Ii represent the interaction cost for action ai given the observed
state of the other agent
•  Represents whether the other agent is along the corresponding path
Communication-less Motion Coordination

[Kimmel,  Bekris  DARS  2014]  


Deconfliction for First-Order Systems

[Pallo$no,  Scordio,  Frazzoli,  Bicchi  ‘07]  


Deconfliction for First-order Systems

[Kron$ris,  Bekris  IROS  ’11]  


Safety Concerns (ICS)
•  Safety becomes a concern in decentralized planning
–  Independently plan paths that are pairwise collision-free
•  For systems with dynamics, e.g., inertia
–  Also avoid inevitable collision states

•  How can communication help?


–  i.e., couple choices in terms of safety considerations
Safe Multi-Robot Motion Coordination

Goal  VB   [Bekris,  Tsianos,  Kavraki  ’07,’09]  

C   current  con$ngency  
 for  C  

plan  A1  
states  
x(tN+2)  

Ini$al  state   Goal  VA  


x(tN+1)  

A   plan  A2  

plan  A3  
current  con$ngency  
 for  B   Goal  VC  
Safe Multi-Robot Motion Coordination

Goal  VB   [Bekris,  Tsianos,  Kavraki  ’07,’09]  

C  

plan  A1  
Ini$al  state   Goal  VA  
x(tN+1)  

A  
plan  A3  
plan  A2  

Goal  VC  
Safe Multi-Robot Motion Coordination

Goal  VB   [Bekris,  Tsianos,  Kavraki  ’07,’09]  

C  

Ini$al  state   Goal  VA  


x(tN+1)  

A  

Goal  VC  
Safe Multi-Robot Motion Coordination

Goal  VB   [Bekris,  Tsianos,  Kavraki  ’07,’09]  

C  

Ini$al  state   Goal  VA  


x(tN+1)  

A  

Goal  VC  
Example
Coordination
If  the  requirements  are  sa$sfied:            Safety  is  guaranteed  
 
How  can  we  implement  the  requirements  for  coordina$on?  
 
Worst-­‐case  
communica$on     Alterna$ve  solu$ons:    
dependency   1.   Global  priority  scheme    
 
 
Problem:  Low  priority  vehicles  do  not    
have  $me  to  compute  a  solu$on    
Effect:  Vehicles  result  oaen  in    
con$ngency  plans    
 
2.  Coopera=ve  Ac=on  Selec=on    
 
Can  the  planning  framework  be  integrated  with  a  balanced,  scalable  
coordina$on  scheme  and  guarantee  safety?  
 
[Bekris,  Tsianos,  Kavraki    ROBOCOMM  ’07]    Best  Student  Paper  Award  
Selection of Contingencies

•  Problem  of  priori$es:   Worst-­‐case  


communica$on    
dependency  
Frequent  selec$on  of  
con$ngency  plans  

•  Casted  the  problem  as  Distributed  Constrained  Op$miza$on  and  


used  a  message-­‐passing  algorithm  (belief  propaga$on  based)  

       #  Vehicles  
Example
Asynchronous Operation
[Bekris,  Grady,  Moll,  Kavraki  -­‐  IJRR  ’12]  

tjn-­‐1   tjn  
Rj   t  
Transmit  πj   Transmit    
ε πi*  
Ri   t  
Plan  πi  
tin-­‐1   tin  
•  Safety challenge: •  Challenges  vs  synchronous  
–  Guarantee that there is a safe path opera$on:  
πi* to select in every planning cycle –  States  cannot  be  accompanied  
by  $mestamps  
–  No  guarantee  messages  arrive  
in  order  
Motion Planning Approaches
Proper$es   •  Classical  approaches  
•  Subop=mal  but  complete  solu=ons  that  are  tractable  
•  New  op=mal  planners  (opportunis=c  decoupling)  
•  Exci=ng  applica=ons:  manipula=on  
Op$mality  
Coupled     How  efficient  solu=ons  
vs   with  formal  guarantees  
can  we  achieve  with  limited  
Decoupled   informa=on  requirements?  
Completeness  
•  Is  it  safe?  
•  Does  it  avoid  deadlocks/
livelocks?  
Safety   •  What  are  its  informa$on  
requirements?  

Approaches  
Centralized   Decentralized  
http://www.pracsyslab.org
Ryan
Luna
Now @
Thank you
Rice Univ. for your
Andrew attention!
Kimmel
•  Push  and  Swap  approach   •  Communica$on-­‐less          
Mo$on  Coordina$on  
•  Dual-­‐arm  scheduling  

Primary  Contributors  

•  Deconflic$on  approach  
•  Pebble  graph  solvers     Athanasios
•  Manipula$on  applica$ons   Krontiris
 
Our research efforts have been supported by:
•  the National Science Foundation (NSF),
•  the National Aeronautics and Space Administration (NASA),
•  the Department of Defense (ONR & DoD TARDEC),

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy