Dexterous  Manipulation Using Hierarchical Reinforcement Learning
Заголовок :
Dexterous Manipulation Using Hierarchical Reinforcement Learning
Подзаголовок:
LearningGripper
Описание:

On a novel pneumatic four-finger gripper with three degrees of freedom per finger we apply reinforcement learning to learn dexter- ous manipulation of objects. In order to reduce the search space, we implemented hierarchical learning on two levels. Low-level learning is used for basic movement primitives like grabbing, lifting or rotation of an object around three cartesian axes, whereas in high-level learning we use the already learned low-level actions to find a policy that ena- bles the gripper to move a target point on the surface of a sphere to the top position in a few seconds. It turns out that Q-learning with a finite state- and action space solves the learning task very well.

Тип документа:
Специальное издание
Заголовок Версия
Dexterous  Manipulation Using Hierarchical Reinforcement Learning
Dexterous Manipulation Using Hierarchical Reinforcement Learning