Enhancing the Path Accuracy of an Industrial Robot with Visual Servoing and Reinforcement Learning

Industrial robots generally exhibit poor path accuracy and thus cannot satisfy many manufacturing requirements. Improving robot path accuracy necessitates motion control with feedback from high-precision external sensors. Typically, feedback control strategies for an industrial robot are realized by adjusting the robot’s intrinsic motion via an external controller. In this research, the control scheme consists of position-based visual servoing (PBVS). An external pose correction controller is introduced to enhance a robot’s path accuracy by reducing the path error. A proportional-integral-derivative (PID) controller is utilized as the primary control method. Due to the repetitiveness of robotic tasks, the control performance can undergo iterative improvement by supplementing the baseline PID controller with a reinforcement learning (RL) based controller, trained via an actor-critic algorithm. The experimental platform comprises two commercial systems: the C-Track 780 dual camera sensor (Creaform) and the M-20iA robot (FANUC). The control performance is assessed using mean absolute error (MAE) and maximum error. The effect of supplementing the baseline PID controller with the proposed RL-based controller is assessed with two experiments. Experiment 1 consists of position-only line following, while experiment 2 consists of full pose line following. Qualitatively, transient performance is improved by overshoot attenuation. In experiment 1, the RL-based controller reduces MAE and maximum errors by 10% and 20%, respectively in the transient phase. The resulting position path accuracy is ±0.09 mm. In the (full pose) experiment, the MAE is not significantly affected, where the maximum errors deteriorate by 4% in position but improve by 9% in orientation. The resulting path accuracy is ±0.30 mm / ±0.07 ◦