Optimizing Non-diagonal Stiffness Matrix of Compliance Control for Robotic Assembly Using Deep Reinforcement Learning

Abstract

Abstract With the advent of Industrial 4.0, industrial robots have been widely used in various sectors, e.g., in assembly. Peg-in-hole (PiH) assembly is the most typical one among assembly tasks. In PiH tasks, the robot transfers from a non-contact mode to a contact-rich mode. Instead of switching the position/force control mode with a pause when contact is detected, we deploy a non-diagonal stiffness compliance control for planning the adaptive trajectory to improve the task efficiency and ensure contact safety. In this paper, we proposed a deep reinforcement learning (DRL) method to achieve the above compliance. A compliance controller based on a virtual forward dynamic (FD) model is built. A DRL agent is deployed to optimize the parameters of the non-diagonal stiffness matrix for the built compliance controller to generate a trajectory that adapts to the changing contact condition. The experiment shows the proposed method can control the contact force within a safe range and improve the efficiency of assembly tasks.