Abstract of FKI-195-94
Title: Supervised Subgoal Generation for Manipulators
Authors: Martin Eldracher, Boris Baginski
Category: Technical Report (Forschungsberichte Künstliche Intelligenz)
Abstract: Building a model for an environment with a specific
manipulator takes exponential computational costs in the
dimension of the manipulator's configuration space.
Furthermore complexity increases with the number of obstacles,
which in real world applications usually is high.
Therefore many classical trajectory planning algorithms,
based on world models, can not cope with a changing
environment. In order to plan complex trajectories, a system
that plans hierarchically shows many advantages. The single
sub-trajectories may be simple and can often be recombined
for new tasks without further low-level planning. In this
article we report on results with different neural network
based (and therefore inherently adaptive), hierarchical
trajectory planning systems. Trajectories are built in
combining known sub-trajectories by choosing subgoals.
The neural systems are trained with the (supervised) back-
propagation learning rule. Nevertheless the subgoals are
produced by the system itself, without any pre-specifications.
We show that useful subgoals can be produced for manipulators
in different (but still static) environments with obstacles.
Opposite to many classical approaches our approach works (once
trained) fast but remains adaptive. In order to show the
capability of our system we compare the results to
a recently introduced model free stochastic search technique.
Keywords: neural networks, supervised learning, subgoal generation,
path planning, manipulator, configuration space, hierarchical
Size: 22 pages
Copyright: The ``Forschungsberichte Künstliche Intelligenz''
series includes primarily preliminary publications,
specialized partial results, and supplementary
material. In the interest of a subsequent final
publication these reports should not be copied. All
rights and the responsability for the contents of the
report are with the authors, which would appreciate