Probabilistic Roadmap Methods (PRMs) solve
the motion planning problem in two phases by sampling free
configurations and connecting them together to build a map
that is used to find a valid path. Existing algorithms are
highly sensitive to the topology of the problem, and their
efficiency depends on applying them to a compatible problem.
Reinforcement learning has been applied to motion planning
and rewards the action performed by planners during either
sampling or connection, but not both.
Previous work computed a global reward and action scheme,
which saw a setback when heterogeneous environments were
concerned. Local learning (connection) was recently introduced
to offset this weakness identified during global learning, and
there was some improvement in planner performance. These
different learning schemes (global and local) have shown
strengths and weaknesses individually.
In this paper, we investigate local learning for sampling.
We study what type of learning to apply when, and how the
two phases of PRM roadmap construction interact, which has
not been investigated before. We show the performance using
each scheme on a KUKAyouBot, an 8 degree of freedom robot,
and analyze what happens when they are all combined during