Exact Collision Checking of Robot Paths - PowerPoint PPT Presentation

About This Presentation
Title:

Exact Collision Checking of Robot Paths

Description:

and connections can be efficiently tested for collision. ... Additional geometric models were provided by PSA (Peugeot-Citroen) Greedy Computation of hij(q) ... – PowerPoint PPT presentation

Number of Views:44
Avg rating:3.0/5.0
Slides: 26
Provided by: lato
Category:

less

Transcript and Presenter's Notes

Title: Exact Collision Checking of Robot Paths


1
Exact Collision Checking of Robot Paths
  • Fabian Schwarzer Mitul SahaJean-Claude Latombe
  • Computer Science Department
  • Stanford University

2
Motivation (1)
One tenet of PRM planning is that sampled
configurations and connections can be
efficiently tested for collision.
3
Motivation (2)
Static collision tests (for sampled
configurations) are done efficiently using
pre-computed bounding volumes (BV) hierarchies
4
Motivation (3)
But dynamic collision tests (for connections)
using BV hierarchies are usually approximate.
  • e too large ? collisions are missed
  • e too small ? slow test of local paths

5
Motivation (3)
But dynamic collision tests (for connections)
using BV hierarchies are usually approximate.
6
Previous Approaches to Dynamic Collision Testing
  • Bounding-volume (BV) hierarchies? Discretization
    issue
  • Feature-tracking methods Lin, Canny,
    91 Mirtich, 98 V-Clip Cohen, Lin, Manocha,
    Ponamgi, 95 I-Collide Basch, Guibas,
    Hershberger, 97 KDS? Geometric complexity issue
    with highly non-convex objects
  • Swept-volume intersection Cameron, 85 Foisy,
    Hayward, 93? Swept-volumes are expensive to
    compute. Too much data.

7
Our Approach
  • ProblemGiven two configurations, test if the
    straight path between them is collision-free, or
    not.
  • Ideas
  • Relate configuration changes to path lengths in
    workspace
  • Use distance computation rather than pure
    collision checking

8
  • Ideas
  • Relate configuration changes to path lengths in
    workspace
  • Use distance computation rather than pure
    collision checking

q (q1,q2,q3) q (q1,q2,q3) dqi qi-qi
For any q and q no robot point traces a path
longer than l(q,q) 3dq12dq2dq3
9
  • Ideas
  • Relate configuration changes to path lengths in
    workspace
  • Use distance computation rather than pure
    collision checking

h(q) Euclidean distance between
robot and obstacles (or lower bound)
h(q)
If l(q,q) lt h(q) h(q) then the straight path
betweenq and q is collision-free
10
  • Ideas
  • Relate configuration changes to path lengths in
    workspace
  • Use distance computation rather than pure
    collision checking

l(q,q) lt h(q) h(q)
11
  • Ideas
  • Relate configuration changes to path lengths in
    workspace
  • Use distance computation rather than pure
    collision checking

l(q,q) gt h(q) h(q)
Bisection
q
q
q l(q,q) lt h(q)
q l(q,q) lt h(q)
12
Greedy Distance Computation
  • Use BV hierarchy same recursion as for pure
    collision checking
  • But compute distance between BVs instead of
    testing BV overlap
  • ? BVs are RSSs
  • returns lower bounds on distance that are
    often much larger than ½ actual distances
  • small factor slower than a pure collision
    checking
  • much faster than BV-based exact or approximate
    distance computation

13
Generalization
  • Robot(s) and static obstacles treated as
    collection of rigid bodies A1, , An.
  • li(q,q) upper bound on length of curve segment
    traced by any point on Ai when robot system is
    linearly interpolated between q and q

l1(q,q) dq1 l2(q,q) 2dq1dq2
l3(q,q) 3dq12dq2dq3
14
Generalization
  • Robot(s) and static obstacles treated as
    collection of rigid bodies A1, , An.
  • li(q,q) upper bound on length of curve segment
    traced by any point on Ai when robot system is
    linearly interpolated between q and q
  • If li(q,q) lj(q,q) lt hij(q) hij(q)then
    Ai and Aj do not collide between q and q

15
Generalized Bisection Method
  • Each pair of bodies is checked independently of
    the others ? priority queue Q of elements
    qa,qbij
  • Initially, Q consists of q,qij for all pairs
    of bodies Ai and Aj that need to be tested.
  • Until Q is not empty do
  • qa,qbij ? remove-first(Q)
  • If li(qa,qb) lj(qa,qb) ? hij(qa) hij(qb) then
  • qmid ? (qaqb)/2
  • If hij(qmid) 0 then return collision
  • Else insert qa,qmidij and qmid,qbij into Q
  • Return no collision

16
Heuristic Ordering Q
  • Goal Discover collision quicker if there is one.
  • Sort Q by decreasing values of
    li(qa,qb) lj(qa,qb) hij(qa) hij(qb)
  • Possible extension to multi-segment paths(very
    useful with lazy collision-checking PRM)

17
Segment Covering Strategies
Allows caching of forward kinematic results
18
Two Arms and Three Rings
Two 20-dof linkages, with 320 triangles
eachThree rings with 6,300 triangles each
19
Robot in a Cage
Robot 2,991 trianglesCage 432 triangles
20
Spot Welding
Robot 2,991 triangles Obstacles 74,681 triangles
21
Comparative Experiment
  • SBL PRM planner (single-query,
    bi-directional, lazy in cc) with
    fixed-discretization collision checker
  • A-SBL Same planner, with new collision
    checkerExperiment
  • Run SBL 10 times on same planning problem with
    some resolution e
  • If a collision has been missed, reduce e and
    repeat
  • If no collision has been missed, return average
    planning time
  • Run A-SBL 10 times and return average planning
    time

Robot 2,502 triangles Obstacles 432
Triangles SBL ? 17 sec A-SBL ? 4.8 sec
22
Some Results
Robot 2,502 triangles Obstacles 432
Triangles SBL ? 17 sec A-SBL ? 4.8 sec
Robot 2,991 triangles Obstacles 432
Triangles SBL ? 83 sec A-SBL ? 44 sec
Robot 2,991 trianglesObstacles 74,681
triangles SBL ? 1.20 sec A-SBL ? 0.81 sec
Robot 2,502 trianglesObstacles 34,171
triangles SBL ? 3.2 sec A-SBL ? 2.1 sec
Robots 6 x 2,991 trianglesObstacles 19,668
triangles SBL ? 85 sec A-SBL ? 52 sec
23
Conclusion
  • New collision checker suited for PRM planners
  • Faster than fixed-resolution checkers
  • Fully reliable
  • Future work
  • Automatic computation of tight upper bounds on
    path lengths in w-space from robot kinematics
  • Better treatment of pairs of moving bodies (e.g.,
    bodies moving along parallel paths)

24
Acknowledgements
  • This research was partially funded by grants from
    General Motors and ABB
  • Additional geometric models were provided by PSA
    (Peugeot-Citroen)

25
Greedy Computation of hij(q)
Idea BV hierarchy same recursion as for pure
collision checking, but compute distance
between boxes (? BVs are RSSs)
  • Algorithm GREEDY-DIST(Ba,Bb)
  • h ? distance(Ba,Bb)
  • If Ba and Bb are both triangles then return h
  • If h gt 0 then return h
  • If Ba is bigger than Bb then switch Ba and Bb
  • a ? GREEDY-DIST(Ba,Bb1)
  • If a gt 0 then
  • b ? GREEDY-DIST(Ba,Bb2)
  • If b gt 0 then return mina,b
  • Return 0

Ba
Bb
hij(q)
  • GREEDY-DIST
  • is small factor slower than a pure collision
    checker
  • is much faster than BV-based exact or
    approximate distance computation
  • returns lower-bounds that are often much larger
    than ½ actual distances
Write a Comment
User Comments (0)
About PowerShow.com