A deep reinforcement learning robot arm control method for balance maintenance of multiple objects on a tray

Abstract

This study aims to develop an optimal control method for a robot arm to maintain the balance of multiple objects on a tray, taking into account the inertia of its movement. As the demand for robot-assisted object transportation increases, the number of objects that conventional robots can grasp remains limited. To solve this problem, a tray is attached to the robot hand, but maintaining the balance of these objects becomes more difficult as the number of joints increases. This study uses reinforcement learning to obtain a robot arm control with the goal of achieving maximum reward when the position, orientation, and velocity of the objects match their initial states without solving complex state equations. This study evaluates the balance control of the robotic arm in a simulation environment, achieving greater performance compared to a static robotic arm, and assesses its potential for implementation in real-world applications.