Motion Planning for Industrial Robots using Reinforcement Learning

A major challenge of today's production systems in the context of Industry 4.0 and Cyber-Physical Production Systems is to be flexible and adaptive whilst being robust and economically efficient. Specifically, the implementation of motion planning processes for industrial robots need to be refined concerning their variability of the motion task and the ability to adaptively deal with variations in the environment. In this paper, we propose a reinforcement learning (RL) based, cognition-enhanced six-axis industrial robot for complex motion planning along continuous trajectories as e.g. needed for welding, gluing or cutting processes in production. Our prototype demonstrator is inspired by the classic wire loop game which involves guiding a metal loop along the path of a curved wire from start to finish while avoiding any contact between the wire and the loop. Our work shows that the RL-agent is capable of learning how to control the robot to successfully play the wire loop game without the need of modeling the wire or programming the robot motion beforehand. Furthermore, the extension of the system by a visual sensor (a camera) allows the agent to sufficiently generalize the learning problem so that it can solve new or reshaped wires without the need of additional learning. We conclude that the applicability of RL for industrial robots and production systems in general provides vast and unexplored potential for processes that feature variability to some extent and thus require a general and robust approach for process automation.