Path Planning Using Lazy PRM

1
Path Planning Using Lazy PRM

• Robert Bohlin, Lydia E. Kavraki
• ICRA 2000, to appear

2
Single Query Path Planning

• No preprocessing
• Ideally...
• Planning time should relate to the difficulty of
  the task
• Planning time should relate to the desired
  quality of the returned path
• The planner should use information from previous
  queries

3
Does PRM work?

• The classical PRM
• has successfully solved problems with many-dofs,
  but...
• Designed to answer many queries require
  preprocessing ( the learning phase )
• heavily rely on fast collision checking

4
Variations of PRM

• Distribute nodes close to the boundary
• Hsu et al (1998) penetration into obstacles
• OBPRM (Amato) rays from collision points
• Boor et al (1999) distribute points in pairs
• Wilmarth et al (1999) retract to medial axis
• Build trees from both directions
• Hsu et al (1997) expansive space
• LaValle Kuffner (1999) attractor

5
A New Principle

• Why difficult for PRM
• Too much time spent on planning local paths that
  will not appear in the final path
• Lazy PRM
• Minimize collision checks!

6
How Does It Work?
7
Node Enhancement

• Recall distribute points close to the boundary
• Strategy pick those edges across the boundary,
  use the midpoints as seeds

Free
Obstacle
Obstacle
8
A Trivial Example

• 1. Generate nodes randomly

G
S
9
A Trivial Example

• 2. Build the initial roadmap

G
S
10
A Trivial Example

• 3. Find a shortest path

G
S
11
A Trivial Example

• 4. Check collisions

G
S
12
A Trivial Example

• 5. Node Enhancement

G
S
13
A Trivial Example

• 6. Find a new shortest path

G
S
14
A Few Details

• Order of collision checking
• Node first, edges next
• End-nodes first
• Edge checking
• coarse-to-fine
• Avoid node clustering
• use an edge in node enhancement only if both ends
  are not from seeds

15
More Details

• Completeness
• half of the new nodes are uniformly distributed
• Multiple Queries
• Scaling
• weighted Euclidean metric

16
Experimental Results

• Distribution of cost
• Collision checking 80
• Map building 18
• Path searching 2
• 26 of the collision checks are on the final
  path!!
• Vs. a naïve PRM a factor of 100-200

17
Discussion

• Ninit, initial nodes, is a key parameter
• Robustness ( worst-case performance )
• Q is Lazy PRM all we need?