Differentially Constrained Planning Mihail Pivtoraiko 1 Motion Planning The Challenge Reliable Autonomous Robots 2 NavLab 1985 Boss 2007 MER 2004 Crusher 2006 ALV 1988 XUV 1998 ID: 373005
Download Presentation The PPT/PDF document "A Survey of" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
A Survey ofDifferentially Constrained Planning
Mihail Pivtoraiko
1Slide2
Motion Planning
The Challenge:
Reliable Autonomous Robots
2
NavLab
, 1985
Boss, 2007
MER, 2004
Crusher, 2006
ALV, 1988
XUV, 1998
Stanford Cart, 1979Slide3
Agenda3
Deterministic planners
Path smoothing
Control sampling
State sampling
Randomized planners
Probabilistic roadmaps
Rapidly exploring Random TreesSlide4
Motivation4
Local
Global
ALV (Daily et al., 1988)Slide5
Unstructured Environments
5
Local
Global
Structure
imposed
:
Regular, fine grid
Standard search (A*)Slide6
6
Global
?
Unstructured and
Uncertain
Uncertain terrain
Potentially changing
LocalSlide7
7
Global
?
Unstructured and
Uncertain
Unseen obstacles
Detected up close
Invalidate the plan
LocalSlide8
8
Global
?
Unstructured and
Uncertain
Efficient
replanning
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(Simmons et al., 1996)
Gestalt (
Maimone
et al., 2002)
LocalSlide9
9Mobility Constraints
2D global planners lead to nonconvergence in difficult environments
Robot will
fail
to make the turn into the
corridor
Global planner
must
understand the need to swing
wideIssues:Passage missed, orPoint-turn is necessary…
Plan Step n
Plan Step n+1
Plan Step n+2Slide10
In the Field…10
PerceptOR/UPI, 2005Rover Navigation, 2008Slide11
11Mobility Constraints
Problem:
Mal-informed global planner
Vehicle constraints ignored:
Heading
Potential solutions:
Rapidly-Exploring Random Tree (RRT)
[
LaValle
&
Kuffner
, 2001]
PDST-EXPLORE
[Ladd &
Kavraki
, 2004]
Deterministic motion sampling
[
Barraquand
&
Latombe
, 1993]
LaValle
&
Kuffner
, 2001Slide12
Mobility Constraints12
Bruntingthorpe Proving Grounds Leicestershire, UK
April 1999Slide13
Extreme Maneuvering13
Kolter et al., 2010Slide14
Arbitrary…14Slide15
Dynamics Planning15Slide16
Definitions
State space
16
x
,
y
,
z
,
, ,
x
,
y
,
z, , ,
vSlide17
Definitions
State space
Control space
Accelerator
Steering
17
x
,
y
,
z
,
, ,
x ,
y,
z, , ,
vSlide18
DefinitionsState space
Control spaceFeasibilitySatisfaction of differential constraintsGeneral formulation
18
x = f
(
x, u, t
)Slide19
DefinitionsState space
Control spaceFeasibilityPartially-known environmentSampled perception map
19Slide20
DefinitionsState space
Control spaceFeasibilityPartially-known environmentSampled perception mapLocal, changing info
20Slide21
DefinitionsMotion PlanningGiven two states, compute control sequence
QualitiesFeasibilityOptimalityRuntimeCompleteness
21Slide22
DefinitionsMotion PlanningDynamic
ReplanningCapacity to “repair” the planImproves reaction time22Slide23
DefinitionsMotion Planning
Dynamic ReplanningSearch SpaceSet of motion alternativesUnstructured environments SamplingState space
Control space
23Slide24
DefinitionsMotion Planning
Dynamic ReplanningSearch SpaceDeterministic samplingFixed pattern, predictable
“Curse of dimensionality”
24Slide25
DefinitionsSearch space designInput: robot properties
Output: state, control samplingMaximize planner qualities (F, O, R, C)Design principleSampling rule
25Slide26
OutlineIntroduction
Deterministic PlanningHierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMsRRTs
Derandomized
Planners
Some applications
26Slide27
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Blocked motions
Free motions
Perception horizon
27
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide28
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
28
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide29
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
29
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide30
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
30
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide31
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
31
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide32
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
32
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide33
Local/Global
Pros:
Local motion evaluation is fast
In sparse obstacles, works very well
Cons:
Search space differences
33
ALV (Daily et al., 1988)
D*/Smarty (
Stentz
& Hebert, 1994)
Ranger (Kelly, 1995)
Morphin
(
Krotkov
et al., 1996)
Gestalt (Goldberg,
Maimone
&
Matthies
, 2002)Slide34
EgographLocal/Global arrangement
Imposing DiscretizationPre-computed search spaceTree depth: 517 state samples per level7 segments4 velocities
19 curvatures
34
Lacaze
et al., 1998Slide35
OutlineIntroduction
Deterministic PlanningHierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMs
RRTs
Derandomized
Planners
Some applications
35Slide36
Path Post-Processing36
Lamiraux et al., 2002
Laumond
, Jacobs,
Taix
, Murray, 1994
Khatib
,
Jaouni
,
Chatila
,
Laumond
, 1997Slide37
Topological Property37
Slide38
OutlineIntroduction
Deterministic PlanningHierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMs
RRTs
Derandomized
Planners
Some applications
38Slide39
Control Space Sampling
39
Barraquand
&
Latombe
, 1993
Lindemann
&
LaValle
, 2006
Kammel
et al., 2008
Barraquand
&
Latombe
:
3 arcs (+ reverse) at
max
Discontinuous curvature
Cost = number of reversals
Dijkstra’s
searchSlide40
Robot-Fixed Search Space
Moves with the robotDense samplingPositionSymmetric sampling
Heading
Velocity
Steering angle
…
Tree depth
1: Local (arcs) + Global (D*) (
Stentz
& Hebert, 1994)5: Egograph (Lacaze
et al., 1998)∞: Barraquand & Latombe
(1993)40
40Slide41
OutlineIntroduction
Deterministic PlanningHierarchicalPath SmoothingControl SamplingState SamplingRandomized Planning
PRMs
RRTs
Derandomized
Planners
Some applications
41Slide42
World-Fixed Search Space
42
Fixed to the world
Dense sampling
(none)
Symmetric sampling
Position
Heading
Velocity
Steering angle…
DependencyBoundary value problem
Pivtoraiko
& Kelly, 2005
Examples of BVP solvers
: Dubins
, 1957 Reeds & Shepp, 1990
Lamiraux & Laumond
, 2001 Kelly & Nagy, 2002 Pancanti et al., 2004
Kelly & Howard, 2005Slide43
Robot-Fixed vs. World-Fixed43
Barraquand & Latombe
CONTROL
STATE
CONTROL
STATE
State LatticeSlide44
State Lattice Benefits44
State LatticeRegularity in state samplingPosition invariance
Pivtoraiko
& Kelly, 2005Slide45
Path Swaths45
Pivtoraiko
& Kelly, 2007
State Lattice
Regularity
Position invariance
Benefits
Pre-computing path swathsSlide46
World Fixed State Lattice46
HLUT
Pivtoraiko
& Kelly, 2005
Knepper
& Kelly, 2006
State Lattice
Regularity
Position invariance
Benefits
Pre-computing path swaths
Pre-computing heuristicsSlide47
World Fixed State Lattice47
?
State Lattice
Regularity
Position invariance
Benefits
Pre-computing path swaths
Pre-computing heuristics
Dynamic
replanningSlide48
World Fixed State Lattice48
State Lattice
Regularity
Position invariance
Benefits
Pre-computing path swaths
Pre-computing heuristics
Dynamic
replanningSlide49
Nonholonomic D*
Expanded States
Motion Plan
Perception Horizon
Graphics: Thomas Howard
49Slide50
Nonholonomic D*50
Pivtoraiko & Kelly, 2007Graphics: Thomas HowardSlide51
Boss“Parking lot” plannerRegular 4D state sampling
Pre-computed search spaceDepth: unlimitedMulti-resolution32 (16) headings2 velocitiesNo curvature
51
LIkhachev
et al., 2008Slide52
World Fixed State LatticeState LatticeRegularity
Position invarianceBenefitsPre-computing path swathsPre-computing heuristicsDynamic replanning
Dynamic search space
52
G
0
G
1
G
3
G
4
G
5
Search graph
G
0
G
1
…
G
n
Pivtoraiko
& Kelly, 2008Slide53
Dynamic Search Space53
Pivtoraiko & Kelly, 2008Graphics: Thomas HowardSlide54
Dynamic Search Space54Slide55
World Fixed State Lattice
55
START
GOAL
State Lattice
Regularity
Position invariance
Benefits
Pre-computing path swaths
Pre-computing heuristics
Dynamic
replanning
Dynamic search space
Parallelized search
Pivtoraiko
& Kelly, 2010Slide56
World Fixed State Lattice
56
START
GOAL
Pivtoraiko
& Kelly, 2010Slide57
START
GOAL
World Fixed State Lattice
57
Pivtoraiko
& Kelly, 2010
epsilon
tree size ratioSlide58
Search Space Comparison58
Robot-FixedPros: Any motion generation schemeCons:
NO Pre-computing path swaths
NO Pre-computing heuristics
NO Parallelized search
NO Dynamic
replanning
NO Dynamic search space
World-Fixed
Pros:
Pre-computing path swaths Pre-computing heuristics
Parallelized search Dynamic replanning Dynamic search space
Cons: Boundary value problemSlide59
OutlineIntroductionDeterministic Planning
HierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMsRRTs
Derandomized
Planners
Some applications
59Slide60
A Few Randomized PlannersProbabilistic Roadmaps (PRM)Kavraki
, Svestka, Latombe & Overmars, 1996Expansive
Space Tree (EST)
Hsu,
Kindel
,
Latombe
& Rock, 2001
Rapidly-Exploring Random Tree (RRT)
LaValle & Kuffner, 2001R* Search
Likhachev & Stentz, 2008
60
LaValle
&
Kuffner, 2001Slide61
Probabilistic RoadmapStatic workspacesE.g. industrial workcells
Two phases:Learning: construct the roadmapQuery: actually planStructure: undirected graph
Originally applied to
holonomic
robots
61Slide62
Learning PhaseTwo steps:ConstructionConstructs edges and vertices to cover free
C-space uniformlyExpansionTries to detect “difficult” regions and samples them more densely62Slide63
Construction StepTwo sub-steps:Sample a random configuration and add to the graph
Select n neighbor vertices and (try to) connect to the new vertexComponentsDistance metricLocal plannerConnections between vertices
63
Kavraki
,
Svestka
,
Latombe
&
Overmars
, 1996Slide64
Query PhaseGraph is constructedApply any shortest-path graph search Smoothing:
Find “shortcuts”Local planner is reused64
Kavraki
,
Svestka
,
Latombe
&
Overmars
, 1996Slide65
OutlineIntroductionDeterministic Planning
HierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMs
RRTs
Derandomized
Planners
Some applications
65Slide66
Original RRTRapidly-Exploring Random TreeProposed for kino
-dynamic planningHolonomic randomized planners existed beforeProposed to meet the need for randomized planners under differential constraintsToday, a work-horse for randomized searchProbabilistically complete
No optimality guarantees
Avoids “curse of dimensionality”Slide67
RRT in a NutshellGiven a tree (initially only x_{init})Pick a sample, x, in state space, X
RandomlySample uniform distribution over XFind nearest neighbor tree node, x_{near}, to xFind a control that approaches x_{near}Unless you solve the BVP problem, won’t approach exactlyJust do your best; you’ll arrive at x_{new} near to x
Add x_{new} and the edge (x_{near}, x_{new}) to the tree
RepeatSlide68
RRT OriginsKhatib, IJRR
5(1), 1986Potential fields for obstacle avoidance, mobile robots and manipulatorsBarraquand & Latombe, IJRR 10(6), 1993
Builds and searches a graph connecting local minima of a potential field
Monte-Carlo technique to escape local minima via
Brownian
motions
Kavraki
,
Svestka
, Latombe & Overmars, Transactions 12(4), 1996Probabilistic roadmaps
Randomly generate a graph in a configuration spaceMulti-query method, best for fixed manipulators
Hsu, Latombe & Motwani, Int. J. of
Comput. Geometry & App., 1997
Expansive Space Tree (EST), single-query methodChoose tree node to extend via biased probability measure
Apply random controlCollision checking in state*time spaceSlide69
Rapidly-ExploringVoronoi bias
Sampling uniform distribution over state spaceLarge empty regions have higher probability of being sampledHence, tree “prefers” growing into empty regions
Naïve random tree
RRTSlide70
Voronoi BiasSlide71
AnalysisConvergence to solutionProbability of failure (to find solution)
decreases exponentially with the number of iterationsProbabilistic completenessSlide72
ExamplesSlide73
Quick overview of CBiRRTConstrained
Bi-directional RRTStart with “unconstrained” RRTAssume some constraints, e.g.:End-effector poseTorque
For each sample, x_{rand}, in X
Get x_{new} by tweaking x_{rand} until constraints satisfied
… via optimization (gradient descent)
In text, “project sample onto constraint manifold”Slide74
Growing the TreeExtending tree toward random sampleThe sample is q_{target}
Nearest neighbor: q_{near}Step from q_{near} to q_{target}In state spaceProject each step onto constraint manifoldUntil you reach q_{target}’s projectionReturn, if can’t continue, e.g. Obstacles
Can’t projectSlide75
Results
5kg
6kg
8kgSlide76
OutlineIntroductionDeterministic Planning
HierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMsRRTs
Derandomized
Planners
Some Applications
76Slide77
Randomness for PlanningPros:Allows rapid exploration of state space
(via uniform sampling)Less susceptible to local minimaCons:Inability to provide performance guaranteesObscures other useful features of plannersMoreover:
There are deterministic incremental sampling methods
E.g.,
Halton
points
Van
der
Corput
sequence (1935); generalized to multiple dimensions by Halton.Slide78
Derandomized RRTLet’s implement
Voronoi bias explicitlyGiven treeCompute Voronoi diagram wrt its nodesPick the sample to extend toward:
Centroid
of largest
Voronoi
region, or
Otherwise reduce size of largest empty ball
Problem:
Voronoi
diagram in arbitrary dimensions – prohibitively expensiveSlide79
Semi-Deterministic RRTSo let’s go for a middle groundInstead of a single x_{rand}Draw a set
k samplesMulti-Sample RRT (MS-RRT)Sort tree nodes acc. to:how many samples they’re nearest neighbor for
Pick node that “collected” most neighbors
Grow tree towards
average
of the neighbor samples
It’s an estimate of the
Voronoi
centroidAs k∞, we get exact Voronoi
centroid“… I shall call him MS-
RRTa!”Slide80
Now, a Deterministic RRT… but with approximate Voronoi bias
Recall: picking k pointsInstead of randomly,Use k uniformly distributed, incremental deterministic samplesE.g., Halton points
The rest stays essentially the same
Meet MS-
RRTb
!Slide81
ResultsLocal minimaBoth MS-RRT are more greedily Voronoi
-biasedLocal minima issues – more pronounced than RRTPaper’s workaround:Introduce obstacle nodes in treeIt’s the nodes that land in obstacles
A mechanism “to remember” not to grow the tree there any more
Sensitivity to metrics
Increased for MS-
RRTa,b
,
Since
Voronoi
depends on metricNearest-neighbor computationMore expensive than O(log n)Increased demand for it in MS-RRTa,bSlide82
ResultsSlide83
OutlineIntroductionDeterministic Planning
HierarchicalPath SmoothingControl SamplingState SamplingRandomized PlanningPRMsRRTs
Derandomized
Planners
Some applications
83Slide84
Mobile ManipulationArbitrary mobility constraints
Optimal solutionUp to representationParallelized computationAutomatically designed
84Slide85
Dynamics Planning85Slide86
SummaryDeterministic planningHierarchical
Path SmoothingControl SamplingState Sampling86
Randomized planning
PRMs
RRTs