Transferring Human Manipulation Knowledge to Industrial Robots Using Reinforcement Learning

Nowadays in the context of Industry 4.0, manufacturing companies are faced by increasing global competition and challenges, which requires them to become more flexible and able to adapt fast to rapid market changes. Advanced robot system is an enabler for achieving greater flexibility and adaptability, however, programming such systems also become increasingly more complex. Thus, new methods for programming robot systems and enabling self-learning capabilities to accommodate the natural variation exhibited in real-world tasks are needed. In this paper, we propose a Reinforcement Learning (RL) enabled robot system, which learns task trajectories from human workers. The presented work demonstrates that with minimal human effort, we can transfer manual manipulation tasks in certain domains to a robot system without the requirement for a complicated hardware system model or tedious and complex programming. Furthermore, the robot is able to build upon the learned concepts from the human expert and improve its performance over time. Initially, Q-learning is applied, which has shown very promising results. Preliminary experiments, from a use case in slaughterhouses, demonstrate the viability of the proposed approach. We conclude that the feasibility and applicability of RL for industrial robots and industrial processes, holds and unseen potential, especially for tasks where natural variation is exhibited in either the product or process.