abstract.tex

Manipulation of deformable cloth objects involves simultaneously
identifying both the grasping location and type of clothing before
beginning manipulation. We extend prior work in this area by learning
an adaptive policy that minimizes the number of manipulations
necessary to correctly identify type and grasp location. By adaptively
choosing a sequence of manipulations, distinct clothing can identified
quickly, while similar pieces will need additional manipulations
before correct identification. Our approach uses a model-based
reinforcement learning approach, which learns a policy in a fully
observable approximation to the full POMDP problem. Our results
demonstrate the benefit of the approach in an existing domain.


We consider augmenting prior work for bringing deformable objects into
desired configurations with a reinforcement learning component to
adaptively determine a policy at runtime. By replacing the hand-tuned
physical manipulation of the cloth with a learned sequence of
manipulations, fewer manipulations can be used for distinctive
situations and more manipulations can be used to discriminate
difficult items of clothing.