Multiresolution PRM Project Update

1
Multiresolution PRMProject Update

• Russell Gayle
• November 10th, 2003
• COMP 290-058

2
Project Goals Review

• Create a multiresolution structure for PRM-based
  planner
• Some multiresolution structures like CLODs have
  quicker distance checking
• Since most of the time is spent computing link
  queries, this could speed up PRMs
• Apply to more complex environments
• For now, those with an arrangement of obstacles
  that makes effective planning difficult
• Maybe apply to scenes with high polygon count as
  well

3
Revised Algorithm/Approach

• Preprocessing
• Compute CLODs for each obstacle and possibly the
  robot
• Multiresolution PRM
• Learning stage Build a roadmap based on coarse
  representations of each obstacle. We can pick
  these by bounding the maximum allowable error for
  distance queries.

4
Algorithm/Approach contd

• Multiresolution PRM
• Query Stage (This is similar to the standard PRM
  query stage)
• Move the robot to the milestone nearest to the
  start point (ensure this is obstacle free)
• Perform a graph search to see if it can reach the
  milestone nearest to the goal
• If such a path exists, move the robot along the
  path until we reach the milestone nearest to the
  goal
• Move the robot from that milestone to the goal
  (ensure this is obstacle free)
• If at any time, the path or milestone moves
  within the error bound, perform a path refining
  step

5
Refinements

• An idea
• Resample with some set number of milestones using
  a finer resolution of the obstacle causing the
  possible error
• Connect these with the milestone before and after
  the place in error and replan locally
• If no path exists locally (in the event that the
  coarser LOD had opened a passage too large, for
  instance), we must replan the entire path

6
Some Problems

• Narrow passages
• Depending on how we refine our meshes, we may
  reduce the effects of narrow passages
• Refinements that are bounded by their parents
  volume or more approximate ones
• Similarly, we could increase the effects of even
  close narrow passages
• Refinements that are more conservative, it has a
  volume greater than that of its finer levels
  (However, any path found in this manner would
  definitely work on finer levels)
• Many refinement algorithms can do both depending
  on the geometry

7
Main work done so far

• Got a well structured PRM planner to work
• Started experimenting with planning using meshes
  at different resolutions
• Currently using GAPS for mesh simplification, but
  many other possibilities, like QSlim, exist
• Learned about the MSL library structure to make
  useful additions
• MSL has a structure for working with different
  types of problems and planners as well as a GL
  renderer and GUI in place
• Includes base classes of Models, Geometry,
  Problem, Solver, and other graphics and gui
  related classes

8
Some results
Based on model from MSL
9
Some results
10
Some results
11
Some results
12
What remains to be done

• Integrate CLOD code with existing planner
• Define a bounded error metric on distance queries
• Allow for quick query-stage path refinements
• Show current LOD for each obstacle
• Test this on many different types of environments