Expert Initialized Reinforcement Learning with Application to Robotic Assembly

This paper investigates the advantages and boundaries of actor-critic reinforcement learning algorithms in an industrial setting. We compare and discuss Cycle of Learning, Deep Deterministic Policy Gradient and Twin Delayed Deep Deterministic Policy Gradient with respect to performance in simulation as well as on a real robot setup. Furthermore, it emphasizes the importance and potential of combining demonstrated expert behavior with the actor-critic reinforcement learning setting while using it with an admittance controller to solve an industrial assembly task. Cycle of Learning and Twin Delayed Deep Deterministic Policy Gradient showed to be equally usable in simulation, while Cycle of Learning proved to be best on a real world application due to the behavior cloning loss that enables the agent to learn rapidly. The results also demonstrated that it is a necessity to incorporate an admittance controller in order to transfer the learned behavior to a real robot.