Machine learning control for human-robot collaboration

Industry 4.0; the fourth industrial revolution is coming and human-robot collaboration (HRC) is one of its decisive principles where human and machine work hand in hand. Both contribute their specific capabilities; human operator controls and monitors production, the robots perform the onerous tasks. In this context, wearable robotics and standard industrial manipulators are common solutions adopted to empower humans. This work focuses on one of the examples of HRC, human-robot co-manipulation. One of the current interest of researchers in this field is to develop an intuitive, and safe control algorithm to minimize the human effort in co-manipulation tasks. With this aim this work presents a variable impedance control and redundancy optimization algorithm for a redundant robotic manipulator arm (KUKA LBR iiwa 14 R820) using the state of the art Model-Based Reinforcement Learning (MBRL) algorithms. An ensemble of neural networks is used to learn the dynamics of the interaction between the human and robot and then it is used with cross-entropy method to tune all impedance control parameters (i.e., stiffness and damping parameters) together with impedance control set-point, online, to minimize human effort. The proposed method has been validated through an experimental procedure and compared with the robot standard impedance controller, showing the capabilities of the designed control strategy, empowering the human operator executing the target task. This work is done in collaboration with CNR-STIIMA as part of H2020 Cleansky 2 EURECA project.