Reinforcement learning control of robot manipulator

Since the establishment of robotics in industrial applications, robot programming involves therepetitive and time-consuming process manually specifying a fixed trajectory, which results machineidle time terms production necessity completely reprogramming for different tasks.The increasing number applications unstructured environments requires not only intelligent butalso reactive controllers, due to unpredictability environment safety measures respectively. This paper presents comparative analysis two classes Reinforcement Learning algorithms, value iteration (Q-Learning/DQN) policy (REINFORCE), applied discretized task positioning robotic manipulator an obstacle-filled simulated environment, with no previous knowledge obstacles’ positions or arm dynamics. The agent’s performance algorithm convergence are analyzed under reward functions on four increasingly complex test projects: 1-Degree Freedom (DOF) robot, 2-DOF Kuka KR16 Industrial random setpoint/obstacle placement. DQN presented significantly better reduced training across all projects third function generated agents both algorithms.