Biasing Samplers to Improve Motion Planning Performance

1
Biasing Samplers to ImproveMotion Planning
Performance

• Shawna Thomas, Marco Morales,Xinyu Tang, and
  Nancy M. Amato
• Parasol Lab, Dept. of Computer Science
• Texas AM University

2
Motivation

• Motion planning has many applications

• Exact motion planning is not practical for many
  problems
• Randomized sampling-based algorithms instead
  trade completeness for computational efficiency

3
Motivation

• Probabilistic Roadmap Methods (PRMs)Kavraki,
  Svestka, Latombe, Overmars, 96

• Roadmap Construction
• Randomly generate robot samples (nodes)
• - discard invalid nodes
• Connect node pairs to form a roadmap
• - simple, deterministic local planner
• - discard invalid paths (edges)
• Query processing
• Connect start and goal to roadmap
• Find path in roadmap between start and goal

Valid
Invalid
C-space
No distribution is the best for all problem
instances
4
Biasing Samplers

• Observation each distribution has its own
  strengths and weaknesses
• Idea we can exploit these strengths by biasing
  one sampling distribution with another

5
Biasing Samplers Framework
Initial problem

Final distribution
6
Experimental Setup

• Goal Compare component sampler combinations in
  different application domains
• Study setup
• Sample 5000 collision free samples with specified
  sampler combination
• Attempt to connect 20 nearest neighbors
• Local planners straight-line and rotate-at-0.5
  Amato et al. 98
• Metrics collected
• Types of samples createdMorales et al. 06
• Diameter changes
• entire roadmap and largest connected component
• Ability to solve a predefined witness query
• samples in narrow passages (when available)

cc-create
cc-merge
cc-expand
cc-oversample
7
Experimental Setup

• Component samplers studied
• OBPRM Amato et al. 98 (OBf and OBc)
• Gauss PRM Boor et al. 99 (Gf and Gc)
• Bridge Test Hsu et al. 03 (BT)
• MAPRM Wilmarth et al. 99 (MA)

8
ResultsRigid Body Problems

• Studied three types of environments

• Trends
• Individual samplers do not in general out-perform
  certain sampler combinations, but best
  combinations come from better components
• Best performing samplers are not the same across
  all environments
• OBfMA performs well in S_Tunnel but not in Walls
• Performance of Gf, Gc, and BT very different in
  S_Tunnel and Walls
• No clear winner in Cluttered

9
(No Transcript)
10
ResultsRigid Body Problems
Narrow passage, thick obstacles

• Roadmap diameters reflect topology evolution
• All samplers quickly map two free portions
• Diameter convergence signals emergence of 1 large
  component
• Diameters for OBfMA grow, converge, and stabilize
  must faster suggesting faster learning
• Witness query solved at different times

Detailed look at runs for 1 random seed
11
ResultsRigid Body Problems
Narrow passage, thick obstacles

• OBfMA boosts performance of MA simply by using
  OBf as starting points
• Exploits OBfs ability to sample closer to
  constrained areas
• Exploits MAs ability to generate samples with
  larger clearances that are easier to connect

Detailed look at runs for 1 random seed
12
ResultsArticulated Linkage Problems

• Studied two types of environments

• Trends
• No samplers solved the witness query (may need
  more sophisticated connection strategies)
• Obstacle thickness affects performance
• OBcMA OBfMA in Hook (colliding surface nodes
  critical)
• OBfMA OBcMA in Maze (free surface nodes
  critical)

13
(No Transcript)
14
Lessons Learned

• By chaining component samplers together, we can
  exploit their individual strengths to create a
  new, better performing, sampling distribution
• Best performing combinations are not the same
  across all environments
• Better combinations usually result from combining
  better components
• Need for strategies to select the appropriate
  sampler combination Vallejo et al. 00,01 Dale,
  Amato 01 Hsu et al. 05 Burns, Brock, 05
• Sophisticated sampling distributions perform
  better when the C-space topology is structured
• In all environments, some combinations generated
  many samples in narrow passages yet still failed
  to solve the query
• Need for more sophisticated connection strategies
  Morales et al. 03, Morales et al. 05

15
Biasing Samplers to ImproveMotion Planning
Performance

• Project Websiteparasol.tamu.edu/groups/amatogrou
  p/research