Adaptive reinforcement learning for opening a door using mobile manipulator in geometrical uncertainty

This paper is a study on an adaptive method for opening a door using a mobile manipulator. In conventional researches on door opening, it has been regarded as a geometry-oriented problem. However, the human does not mind the trajectory but just adapt to the exerting forces, while he/she opens a door. The main idea of this research is that we treat the door opening as a force-oriented problem. We will not assume a door's exact trajectory. Instead of this, only compliance control is executed with a simple command which directs a pulling direction. Finally, the mobile manipulator adapts itself to the door trajectory, with bounded force condition. The main challenge of this research is to find an optimum compliance gain at each configuration of the manipulator, with position error of mobile base. To resolve this problem, we employed an adaptive method based on modern reinforcement learning. Also, we adopt the concept of compliance ellipsoid, which is the graphical representation of a compliance matrix, for the proposed RL algorithm. In this work, we simulated a door opening task, and the simulation results prove that the proposed adaptive strategy was successfully fulfilled the door opening constraints, with permitting large error bound of the mobile base position.