Sampling and Searching Methods for Practical Motion Planning Algorithms
Robot Motion Planning Methods: An...
Transcript of Robot Motion Planning Methods: An...
Robot Motion Planning Methods: An Overview 16 April 2007 1
Robot Motion Planning Methods: AnOverview
Bhaskar Dasgupta
Department of Mechanical EngineeringIndian Institute of Technology Kanpur
http://home.iitk.ac.in/˜ dasgupta
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 2
Outline
1. The Basic Problem
2. Mathematical Formulation
3. Roadmap Methods
4. Cell Decomposition Methods
5. Potential Field Methods
6. Comparison of Methods
7. Extensions of the Basic Problem
8. Summary
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 3
The Basic Problem
We have a rigid object, the robot, of known geometry, capable of freely movingin a workspace that contains a number of fixed rigid objects, calledobstacles, of known geometry and location.
Problem:
Given initial position and orientation and goal position and orientation ofthe robot in the workspace,
generate a path specifying a continuous sequence of positions andorientations of the robot
avoiding contact with the obstacles,
starting at the initial position and orientation and
terminating at the goal position and orientation.
Report failure if no such path exists.
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 4
Mathematical Formulation
• Path Planning and Collision Avoidance → Spatial representation
– Workspace W
– Obstacle space WOi of all obstacles i
– a reference point of the robot
Workspace
Obstacle
Robot
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 5
Def.: Configuration q of a Robot:complete specification of the position of all points of the robot
Def.: Configuration space or C-space Q:set of all possible configurations
• Configuration R(q) in the C-space is represented by a point
• Dimension of the C-space is (usually) the number of degrees of freedom
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 6
• Example: A robot with q1 =
x1
y1
ϕ1
ϕ1y1
x1
y
x
−→
y1
x1
y
x
ϕ1
ϕ
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 7
• Obstacles → Representation in the configuration space
Def.: C-obstacle QOi:Set of configurations, at which the robot collides with (intersects) theobstacle i
QOi = {q ∈ Q |R(q)⋂
WOi 6= ∅}
• Free space: Qfree = Q\ (⋃
i QOi)
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 8
• Example: A robot with translatory motion in the plane
– Workspace and obstacle
Workspace
Obstacle
Robot
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 8
• Example: A robot with translatory motion in the plane
– Configuration space
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 8
• Example: A robot with translatory motion in the plane
– Obstacle in the configuration space
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 8
• Example: A robot with translatory motion in the plane
– Obstacle in the configuration space
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 8
• Example: A robot with translatory motion in the plane
– Free space
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 9
• Path planning: To find a free path for the robot
– from an initial configuration qStart
– to a final configuration qGoal
– in the free space Qfree
• Mathematicallycontinuous mapping c : [0, 1] → Qfree
such that c(0) = qStart and c(1) = qGoal
• Free path: contact with obstacles not allowed
• Semi-free path: contact with obstacles, though not intersection
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 10
Roadmap Method
• Roadmap: a network or graph of connections
• Only prescribed connections between specific pairs of end-points are de-veloped
• Path finding from Start to Goal
– Collision-free path fromStart to Roadmap
– Path along the Roadmapto the neighbourhood of the Goal
– Collision-free path fromthere to Goal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 11
• Representation of the Roadmap as a non-directed graph
– Nodes: Salient positions
– Edges: Paths between neighbouring nodes
Def.: Roadmap RM :Union of line segments, such that for all qStart ∈ Qfree and qGoal ∈
Qfree there exist
• a path from qStart to a q′Start ∈ RM
• a path from a q′Goal ∈ RM to qGoal
• a path from q′Start to q′Goal in RM
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 12
• Shortest path (pathway) can be found by the usual algorithms of graph theory
• Construction of the network
– Visibility graph
– Reduced visibility graph
– Generalized Voronoi diagram
– Silhouette method
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 13
Visibility graph
• Defined for
– two-dimensional configuration space
– polygonal C-obstacles
• Nodes vi of the visibility graph
– qStart and qGoal
– vertices of the C-obstacles
• Edges eij
– connect pairs of nodes, that are ‘visible’ from each other
– a line segment joining two such nodes does not intersect any obstacleeij 6= ∅ ⇐⇒ svi + (1 − s)vj ∈ Qfree ∀s ∈ [0, 1]
– edges of obstacles are also taken as edges of the graph
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 14
Example: Visibility graph
Start
Goal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 15
Reduced visibility graph
• Not all edges of the visibility graph are included
• Only supporting edges and separating edges are used
– Supporting edge: obstacle(s) on the same side
– Separating edge: individual obstacles completely on different sides
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 16
• Example
Start
Goal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 16
• Example: Reduced visibility graph
Start
Goal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 17
Generalized Voronoi diagram (GVD)
• Retraction of Qfree: GVD
• Paths maintain clearance from the obstacles
• Two-dimensional C-space
• GVD V of Qfree
– Set of points from Qfree
– Each point of the set at equaldistance from both of itsnearest C-obstacles
• Every point q ∈ Qfree possesses animage ρ(q) ∈ V
ρ ( )1qρ ( )2q1
q1
q2
v
2e
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 18
Silhouette method
• In n-dimensional C-space
– An (n − 1)-dimensional hyperplane is swept through Qalong the remaining dimension,e.g. in xi-direction, andproduces slices
– Extreme points inxj -direction (xj⊥xi) on theslices are marked
– Set of the marked pointsconstitutes the silhouette curves
– Silhouette curves, in general,not connected
x 2
x 1
Slic
e
Extreme points
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 18
Silhouette method
• In n-dimensional C-space
– An (n − 1)-dimensional hyperplane is swept through Qalong the remaining dimension,e.g. in xi-direction, andproduces slices
– Extreme points inxj -direction (xj⊥xi) on theslices are marked
– Set of the marked pointsconstitutes the silhouette curves
– Silhouette curves, in general,not connected
2
1
x
x
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 19
– Slices, at which the number ofsilhouette curves changes: critical slices∗ new (recursive) application of the algorithm with the slice
as (n − 1)-dimensional C-space∗ (n − 2)-dimensional hyperplane sweeps across the critical slice∗ one-dimensional slice→ slice is the silhouetteand recursion terminates
• qStart and qGoal are consideredas extreme points
2
1
x
x
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 19
– Slices, at which the number ofsilhouette curves changes: critical slices∗ new (recursive) application of the algorithm with the slice
as (n − 1)-dimensional C-space∗ (n − 2)-dimensional hyperplane sweeps across the critical slice∗ one-dimensional slice→ slice is the silhouetteand recursion terminates
• qStart and qGoal are consideredas extreme points
2
1
Start
x
x
Goal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 20
Cell Decomposition Method
• Cell decomposition
• Segmentation of the free space Qfree into cells
• Crossing over possibilities between cells are ascertained based on adjacency
– neighbouring cells: having common boundary
– crossing over only between neighbouring cells
• Representation of the crossing over possibilities as a graph
• Different methods
– Exact cell decomposition
– Approximate cell decomposition
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 21
Exact cell decomposition
• Union of the cells describes the free space Qfree exactly
• Assumptions
– two-dimensional C-space
– polygonal boundary of C-space
– polygonal C-obstacles
• Several possible strategies, e.g.
– Trapezoidal
– Polygonal
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 22
• Example: trapezoidal decomposition, vertical trapeziums
– from all vertices, vertical lines
– example: xy-coordinate system→ Lines parallel to y-axis
– end-points: boundary of C-space or C-obstacle
– enumeration of the cells
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 22
• Example: trapezoidal decomposition, vertical trapeziums
– from all vertices, vertical lines
– example: xy-coordinate system→ Lines parallel to y-axis
– end-points: boundary of C-space or C-obstacle
– enumeration of the cells
c 15
c 2
c 1 c 10
c 14
c 7c 5
c 4
c 3
c 6
c 11
c 13
c 9 c 12
c 8
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 23
– Graph development based on adjacency
– Sequene of cells: channel
– Path finding: line segments between mid-points of the vertical boundari-es
c 15
c 2
c 1 c 10
c 14
c 5c 4
c 3
c 6
c 11
c 13
c 9
c 7
c 12
c 8
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 24
Approximate cell decomposition
• Fundamental segmentation
• Entire free space is segmented into
– simple and
– similar
cells of specified shape
• Cells must be non-overlapping
• Characterization of the cells: covered by C-obstacles
– completely: filled cell
– partly: Mixed cell
– not at all: empty cell
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 25
• Example: identical rectangles as cells
qstart
qgoal
• Only approximate description of the free space possible
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 26
• Maximum error: dependent on the cell size
– too large cell size: bottlenecks may remain unmanoueverable
– too small cell size: huge number of cells, inefficiency
• Hierarchical methods, e.g. Divide-and-label
– Gradual reduction of error
– mixed cells → further subdivision
– Continuation of the method through several levels of hierarchy till the requiredaccuracy is achieved
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 27
• Example: Quadtree
– Begin with root cell, containing the entire free space
– Subdivide the mixed cells hierarchically∗ each dimension of a cell is halved∗ till the required accuracy
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 27
• Example: Quadtree
– Begin with root cell, containing the entire free space
– Subdivide the mixed cells hierarchically∗ each dimension of a cell is halved∗ till the required accuracy
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 27
• Example: Quadtree
– Begin with root cell, containing the entire free space
– Subdivide the mixed cells hierarchically∗ each dimension of a cell is halved∗ till the required accuracy
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 27
• Example: Quadtree
– Begin with root cell, containing the entire free space
– Subdivide the mixed cells hierarchically∗ each dimension of a cell is halved∗ till the required accuracy
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 28
– Efficient storage as a tree: “Quadtree”
∗ Nodes of the tree: cells∗ Root of the tree: the root cell∗ Every mixed cell has four child nodes: NW, NE, SW, SE
– 3-dimensional: Octree
– n-dimensional: 2n-tree
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 29
Potential Field Methods
• Robot (as point in C-space Q) under influence of artificial field forces
– attractive force of the goal
– repulsive force of the C-obstacles
• Artificial forces generated through artificial potential fields
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 30
• C-space described throughpotential function U : Q → R
for every point q ∈ Q
• Example: attractive potential Uatt of the Goal
– Potential strength increases with distance!
– quadratic variation with distance
– Uatt(q) = 1
2ζ ||q − qGoal||
2
– with ζ ∈ R as scale factor
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 31
• Example: repulsive potential Urep of the C-obstacle
– infinite inside the obstacle
– decays fast to zero at a threshold distance d0
– Urep(q) =
1
2η
(
D−1(q) − d−1
0
)2
for D(q) ≤ d0
0 for D(q) > d0
– with D(q) as clearance from nearest obstacle
– with η as scale factor
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 32
• Total potential field: U(q) = Uatt(q) + Urep(q)
• With U differentiable,Force vector F (q) = −∇U(q)
• Robot follows negative gradient −∇U(q)
• Motion of the Robot
– explicit solution of the dynamic equations, or
– incremental: steps in the negative gradient direction→ a real-time process
• Local minima
– navigation around the neighbouring obstacle
– definition of a new potential function
– random deviations
– incorporation of global considerations
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 33
Potential field variants
• Combination with cell decomposition
– operate along pre-planned path based on a priori crude knowledge of theC-space
– take corrective action from potential field based on sensory input
• Off-line applications: global (computation-intensive) considerations
– Variational formulation
– Global navigation function
– Domain mapping
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 34
Probabilistic Roadmaps
• PRM: similar to conventional roadmaps
• Nodes of the graph randomly selected from Qfree
• Every node is connected with “neighbours”
– as long as the connection intersects no obstacle
– number of neighbours: implementation-dependent
• Roadmap can be successivelyrefined/enhanced
• Several strategies to
– locate nodes
– connect them
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 35
Comparison of the Methods
Roadmap Cell decomposition Potential field
VG / GVD / Sil exact / approx local / global
Dimensions 2 / 2 / arbitrary 2 / arbitrary arbitrary
Completeness yes yes / no no / (yes)
Optimality yes / no? / no no no / YES
Complexity strongly dependent on dimension of Q and
obstacles (number, complexity)
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 36
Extensions of the Basic Problem
• Moving obstacles
– a priori knowledge ?
– time as another variable in C-space, or CT-space
– multiple robots
– articulated manipulation robots
• Movable obstacles
– C-space can be modified by the robot
• Kinematic constraints
– Holonomic constraints: reduction of the dimension of Q
– Non-holonomic constraints: car-like robots
• Uncertainities
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 37
Summary
• Motion planning in Robotics
• Classes of methods
– Roadmap methods
– Cell decomposition
– Potential field
• Comparison of methods
• Extensions of the basic problem
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 38
Literature
• Howie Choset, Kevin M. Lynch, Seth Hutchinson, George Kantor, Wolfram Burgard,Lydia E. Kavraki and Sebastian Thrun. Principles of Robot Motion. MIT Press, 2005.
• Jean-Claude Latombe. Robot Motion Planning. Kluwer Academic Publishers, 3.Auflage, 1993.
• S. M. LaValle. Planning Algorithms. Cambridge University Press(http://msl.cs.uiuc.edu/planning/), 2006.
• Gregory Dudek and Michael Jenkin. Computational Principles of Mobile Robotics.Cambridge University Press, 2000.
• Richard M. Murray, Zexiang Li and S. Shankar Sastry. A Mathematical Introduction toRobotic Manipulation. CRC Press, 1994.
• Robert J. Schilling. Fundamentals of Robotics. Prentice Hall, 1990.
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur
Robot Motion Planning Methods: An Overview 16 April 2007 39
Acknowledgements
Thanks to
• audience of my classrooms of2001, 2003, 2003 (summer), 2005, 2005 (Berlin) and 2006
• Dr.-Ing. habil. Dietmar Tutsch of TU Berlin.
Thank you!
Bhaskar Dasgupta INDI
AN
INSTI
TUTE OF TECHN OLO
GY
KANPURIIT Kanpur