Door opening by joining reinforcement learning and intelligent control

In this paper we address a problem of how to open the doors with an articulated robot. We propose a novel algorithm, that combines widely used reinforcement learning approach with intelligent control algorithms. In order to speed up learning, we formed more structured search, which exploits physical constraints of the problem to be solved. The underlying controller, which acts as a policy search agent, generates movements along the admissible directions defined by physical constraints of the task. This way we can efficiently solve many practical problems such as door opening without almost any previous knowledge of the environment. The approach was verified in simulation as well as with real robot experiment.