Reinforcement Learning for Manipulators without Direct Obstacle Perception in Physically Constrained Environments

摘要

Reinforcement Learning algorithms have the downside of potentially dangerous exploration of unknown states, which makes them largely unsuitable for the use on serial manipulators in an industrial setting. In this paper, we make use of a policy search algorithm and provide two extensions that aim to make learning more applicable on robots in industrial environments without the need of complex sensors. They build upon the use of Dynamic Movement Primitives (DMPs) as policy representation. Rather than model explicitly the skills of the robot we describe actions the robot should not try to do. First, we implement potential fields into the DMPs to keep planned movements inside the robot's workspace. Second, we monitor and evaluate the deviation in the DMPs to recognize and learn from collisions. Both extensions are evaluated in a simulation